Category id 1f929d12f45cfa62 (derived from Scout's category name (no id in Scout's data)). 29 alternatives, shown in Scout's order.
Re-run this test
Transcriber head-to-head, 2026-11-18 (16 clips: clean, accented, noisy, non-English, long). Take the public inputs and references, run any system on them, and score it with the same scorer. Our own result is in the table, including where we lose.
Pooled word error rate, all clips run (lower is better)
System
Pooled WER
Clips OK / run
groq turbo
2.05%
16 / 16
steadyfetch
2.42%
16 / 16
groq large-v3
3.23%
16 / 16
tidytools
3.25%
15 / 15
ours_small (ours)
5.40%
15 / 15
memo23_small
10.67%
4 / 4
Systems ran different clip sets (groq turbo 16, steadyfetch 16, groq large-v3 16, tidytools 15, ours_small 15, memo23_small 4), so pooled numbers compare like with like only on shared clips. The full report has per-clip results.
OpenAI Whisper API provides automatic speech recognition (ASR) via REST endpoint, accepting audio files in MP3, MP4, MPEG, MPGA, M4A, WAV, and WEBM formats up to 25MB per request. The API transcribes audio to text with optional output of detected language and word-level timestamp confidence scores. The model is trained on 680,000+ hours of multilingual audio data from the web, enabling accurate transcription... (Scout's description)
CONFIRMED 2026-10-19: whisper-1 is no longer listed on the pricing page. Listed transcription models: gpt-transcribe $0.0045/min; gpt-4o-transcribe $0.006/min; gpt-4o-mini-transcribe $0.003/min; gpt-live-transcribe $0.017/min; gpt-realtime-whisper $0.017/min. whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe and gpt-4o-transcribe-diarize shut down Feb 26, 2027. as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Fastest-to-integrate cloud API; excels at multilingual support and free-tier availability. Limited by 25MB file size per request. Handles technical English moderately; noisy/accented audio less reliable than commercial competitors. Best value for cost-sensitive batch work. [2026-10-19T22:08Z cloud refresh: 2026-10-19 freshness check: OpenAI deprecation page, dated 2026-08-26: 'we notified developers using whisper-1, gpt-4o-transcribe, gpt-4o-mini-transcribe, and gpt-4o-transcribe-diarize of their deprecation and removal from the API on February 26, 2027.' Recommended replacement: gpt-live-transcribe or gpt-transcribe. Stored diarization note referred to gpt-4o-transcribe-diarize, which is also being removed. two_year_cost_gbp left as is (was empty; no usage basis stored).] [2026-10-19T22:08Z cloud refresh: Q4-2026 review: shutdown announced. OpenAI notified on 2026-08-26 that whisper-1, gpt-4o-transcribe and gpt-4o-mini-transcribe are removed on 2027-02-26, with gpt-transcribe or gpt-live-transcribe as replacements. Agent judgement: park this entry (no new builds on it) and evaluate gpt-transcribe, which OpenAI prices at $0.0045/min vs $0.006/min. gpt-transcribe is not yet a DB entry. Pricing for this entry is carried by patches/02-price-refresh.json (fuller text, same facts); apply 02 before 03.] [2026-10-20T01:33Z cloud refresh: type proprietary -> commercial (proprietary; pricing paid; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from OpenAI Services Agreement (no reverse engineering, closed service); api_or_webhooks from speech-to-text guide (now served at developers.openai.com). self_hostable left out: the hosted API has no self-hosted edition, but the open Whisper weights are a separate self-hostable product, so false would mislead. data_export n/a (stateless compute API).]
Groq Whisper (paid tier)
commercialnot runnable on gaents yet
Groq Whisper delivers ultra-fast speech-to-text via proprietary LPU (Language Processing Unit) inference hardware. The service runs OpenAI's Whisper Large v3 and Whisper Large v3 Turbo models at 189x and 228x real-time speed respectively, meaning 1 hour of audio transcribes in ~19 seconds (Large v3) or ~16 seconds (Turbo). Files up to 25MB (free tier) or 100MB (Developer tier) are supported, with 10-second... (Scout's description)
CONFIRMED 2026-10-19: whisper-large-v3 $0.111 per hour of audio; whisper-large-v3-turbo $0.04 per hour; both listed as Production models, max file size 100 MB, Developer-plan limits 200K ASH / 300 RPM (v3) and 400K ASH / 400 RPM (turbo). Minimum billed length 10 seconds. groq.com/pricing redirects to the homepage; prices read from console.groq.com/docs/models. as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Exceptional throughput: 228x real-time for Turbo, 164x for Large v3. Cost 89% cheaper than OpenAI Whisper ($0.04/hr vs $0.36/hr). Free tier modest (8/day), but paid Developer tier reasonable. Accuracy on technical English: ~10.3% WER. Best for high-volume, latency-sensitive batch jobs. File size limit: 25MB free, 100MB Developer tier. [2026-10-19T22:08Z cloud refresh: 2026-10-19 check: Groq's site has been relaunched as an 'AI inference cloud' working alongside NVIDIA GPUs; the public pricing page is gone. Official newsroom (August 17, 2026): 'Groq LLC ("Groq") today announced a $350 million Series A fundraise and the round was led by Disruptive, with planned participation from NVIDIA.' Stored prices ($0.111/h, $0.04/h) kept but now unverified. two_year_cost_gbp unchanged (null).] [2026-10-19T23:00Z cloud refresh: Prices re-confirmed: patch 02 marked them UNVERIFIED because groq.com/pricing redirects to the homepage, but the console models page (console.groq.com/docs/models.md) still lists both Whisper prices. url changed from groq.com/pricing (now a redirect) to the speech-to-text docs. No Whisper deprecation on Groq's deprecations page (the last one is distil-whisper-large-v3-en, 2025-08-23). Verdict evaluate not changed. Note: patch 02 also set pricing for this row; this text supersedes it if both are applied.] [2026-10-20T01:33Z cloud refresh: type proprietary -> commercial (proprietary; pricing paid; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: api_or_webhooks from speech-to-text docs; licence_class left out (only website terms found, not the Services Agreement); self_hostable left out; data_export n/a (inference API).] [2026-10-23T13:15Z cloud refresh: Fact fields pass 2 (2026-10-20): licence_class proprietary from the GroqCloud Services Agreement (covers the speech-to-text API as part of the Cloud Services). self_hostable left out (see Groq row). data_export n/a (transcription API).]
Deepgram Nova-3
commercialnot runnable on gaents yet
Deepgram Nova-3 is a unified multilingual speech-to-text model released February 2025, designed for real-time and batch transcription with emphasis on accuracy, multilingual support, and low latency. The model employs a sophisticated audio embedding framework using representation learning to project audio into a compressed latent space, enabling in-context learning for dynamic vocabulary customization at... (Scout's description)
CONFIRMED 2026-11-09: Nova-3 Monolingual pre-recorded $0.0043/min Pay As You Go ($0.0036/min Growth); Nova-3 Multilingual pre-recorded $0.0052/min ($0.0043/min Growth). Streaming Nova-3 Monolingual $0.0077/min PAYG ($0.0065/min Growth); the earlier $0.0048/min discounted 'current price' is no longer shown. as of 2026-11-09
Scout, alternatives.json v3.72026-11-09date written in Scout's pricing text
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
Best-in-class real-time latency (<300ms p95). Multilingual unified model with code-switching. Keyterm prompting (up to 100 terms) without retraining. Strong on technical terminology via representation learning. Slightly higher cost than Groq but lower than OpenAI. Concurrent connection limits: 50 (REST), 150-225 (WebSocket) depending on plan. · 2026-09-14 PRICE CORRECTION (S38): pre-recorded was on record at $0.0077/min, actually $0.0043/min — 44% stale in the OVER-costing direction. We were pricing an ADOPTED tool too high, which distorts every build-vs-buy that used it as the buy-side benchmark. [2026-10-19T22:08Z cloud refresh: Q4-2026 review: the 2026-09-14 'correction' from $0.0077 to $0.0048 streaming is a discounted current price; the page still lists $0.0077/min as the regular price. Latest changelog entry Oct 12, 2026. Verdict unchanged.] [2026-10-20T01:33Z cloud refresh: type proprietary -> commercial (proprietary; pricing paid; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from Deepgram terms; self_hostable from self-hosted docs (licensed self-hosted deployment exists; not checked whether Nova-3 self-hosting needs an enterprise contract); api_or_webhooks from the same docs (API requests); data_export n/a (docs say request data is not stored).]
AssemblyAI
freemiumnot runnable on gaents yet
AssemblyAI is a cloud-hosted automatic speech recognition platform launched with Universal-3 Pro model in February 2026, emphasizing accuracy, multilingual support, and integrated audio intelligence features. Universal-3 Pro represents a new class of speech language model trained on tens of thousands of hours of specialized audio data with improved Efficient Conformer audio encoder architecture, achieving ~98.4%... (Scout's description)
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Best English accuracy (1.56% WER on LibriSpeech). 57% better recognition of key terms (names, codes, medical terms) vs prior models. Diarization, summarization, sentiment, entity detection built-in (paid add-ons). Streaming now available with real-time features. Note: speaker identification outputs generic labels (Speaker A/B/C), not actual names. Smaller mindshare (5.9% on PeerSpot, down from 9.2%) vs Deepgram. | 2026-10-02 (withheld D554, 09-23, n=1 67.6-min episode, domain-term spelling): Universal-3.5 Pro + keyterms 99% ($0.26/h, ~$653 corpus); Universal-2 + keyterms 96% ($0.15/h, ~$377); Groq Whisper + prompt 48% (~$100); a glossary map lifts Whisper to 97% but ~10% of its edits destroy correct terms. wolfram/requests/from-withheld-2026-09-23-d554-arms-c-d-results.md [2026-10-19T22:08Z cloud refresh: Best English accuracy (1.56% WER on LibriSpeech). 57% better recognition of key terms (names, codes, medical terms) vs prior models. Diarization, summarization, sentiment, entity detection built-in (paid add-ons). Streaming now available with real-time features. Note: speaker identification outputs [2026-10-19 Scout T2 price/licence check: Model renamed: top async model is now 'Universal-3.5 Pro' at the same $0.21/hr. Summarization changed from $0.03/hr to $0.02/hr (low) or $0.07/hr (medium). Free tier now stated as hours (185 h pre-recorded, 333 h streaming), not '$50 free credits'. Streaming now has list prices.]] [2026-10-20T01:33Z cloud refresh: type proprietary -> freemium (proprietary; pricing has free tier + paid; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from Terms of Service; api_or_webhooks from docs; self_hostable left out (self-hosted docs URL tried returned 404); data_export left out.] [2026-10-23T13:15Z cloud refresh: Fact fields pass 2 (2026-10-20): self_hostable true from AssemblyAI's Self-hosted deployment page (models run on your own infrastructure; sales-led). data_export n/a (transcription API).]
Rev.ai
commercialnot runnable on gaents yet
Rev.ai is a speech-to-text API platform with dual-mode transcription: automatic speech recognition via Reverb model and optional human professional transcription. Reverb ASR is a custom model trained on 200,000 hours of human-transcribed English audio—the largest corpus used to train an open-source ASR model. The model achieves 96% accuracy (self-reported) with 9.68% WER on Earnings21, 13.68% on Earnings22, and... (Scout's description)
CONFIRMED 2026-10-19 (Pay As You Go): Reverb Transcription (English) $0.20 per hour; Reverb Foreign Language $0.30 per hour; Whisper Large $0.005 per minute; Human Transcription $1.99 per minute; Forced Alignment $0.003/min. Free credits = 5 hours of Reverb ASR. Enterprise volume pricing. as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Cheapest machine-only option at $0.003–$0.005/min. Reverb (proprietary) achieves 96% accuracy claim (self-reported, not independently benchmarked vs other providers' test sets). Open-source inference code available via GitHub (revdotcom/reverb). Strong on English; single-language-per-job limitation (no mid-call code-switching). Best for cost-sensitive English-only use cases; less suitable for technical English or multilingual workloads. [2026-10-19T23:00Z cloud refresh: 2026-10-19 price refresh: Reverb is $0.20/hour (stored ~$0.18). 'Whisper Fusion' no longer listed; only 'Whisper Large'.] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: api_or_webhooks from API docs (Asynchronous and Streaming Speech-to-Text APIs). licence_class left out (rev.ai/terms 404). self_hostable left out (no deployment page found). data_export n/a (transcription API).]
whisper.cpp
open-sourcenot runnable on gaents yet
whisper.cpp is a high-performance C/C++ port of OpenAI's Whisper automatic speech recognition model, enabling efficient, standalone transcription on consumer hardware without external dependencies, Python runtime, or ML frameworks. The project uses plain C/C++ implementation with zero runtime memory allocations, mixed F16/F32 precision, and integer quantization support for speed and memory efficiency.
Hardware... (Scout's description)
Free (open-source). Hardware cost for GPU acceleration: NVIDIA GPU rental (e.g., L40S at $0.72/hr on-demand, $0.32/hr spot on Spheron) for 60+ concurrent streams. CPU-only: free but ~3–10x slower. Comparative: 2,513 hours at L40S $0.32/hr spot ≈ $24.60 one-time cost if batch-processed at 100x real-time. as of: no date recorded for this field
Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
GitHub stars
55199 as of: no date recorded for this field
Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Self-hosted, no API lock-in, data never leaves your infrastructure. Model sizes: tiny (75MB/273MB RAM) to large (2.9GB/3.9GB RAM). Supports NVIDIA CUDA, Apple Metal, Vulkan, OpenVINO, CPU-only. Accuracy matches or beats commercial APIs on clean audio; handles noisy/accented audio well. Tradeoff: integration effort (8+ hours), hardware capex. Break-even: CPU processing ~$0. GPU rental for 2,513 hours: $24–500 depending on concurrency and spot vs on-demand pricing. [2026-10-23T13:15Z cloud refresh: Sweep 2026-10-20: last commit 2026-10-19; licence unchanged (MIT).]
SearchAPI YouTube Transcripts API
commercialnot runnable on gaents yet
SearchAPI YouTube Transcripts API is a cloud-based service that fetches YouTube video transcripts and subtitles via a simple REST API endpoint, handling the complexity of YouTube's transcript delivery without requiring direct authentication or browser automation. The service manages all proxy rotation and rate-limiting internally, allowing developers to request transcripts by video ID and receive structured data... (Scout's description)
CONFIRMED 2026-10-19: Pay per successful search (failed requests not billed). Developer $40/month for 10,000 searches ($4 per 1,000); Production $100/month for 35,000 ($3/1k); then $250/100k, $500/250k, $900/500k, $1,500/1M, $2,800/2M, $5,000/5M ($1/1k). 100 free requests at signup. as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Vendor-side proxy rotation (100M+ global proxies); bypasses YouTube IP blocks transparently without user configuration; supports both auto-generated and manual captions with language preference; does NOT run STT (fetches YouTube natives only). [2026-10-19T22:08Z cloud refresh: Changed: stored per-request figure was an estimate ('$0.003–0.01 per transcript based on SERP API comparison'). The official pricing page gives $0.004 per search on the $40 Developer plan. The page prices 'searches'; it does not say a transcript call is billed differently, so one transcript is treated as one search (agent assumption). two_year_cost_gbp: agent-chose basis = Developer plan $40 x 24 months = $960 at £0.7502 per $1 = £720.19.] [2026-10-20T01:33Z cloud refresh: type proprietary -> commercial (no recurring free tier: 100 free requests at signup (one-off); rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from SearchApi terms (proprietary technology); api_or_webhooks from API docs; self_hostable left out; data_export n/a (stateless API).] [2026-10-23T13:15Z cloud refresh: Fact fields pass 2 (2026-10-20): self_hostable false: hosted API only (endpoint https://www.searchapi.io/api/v1/search; searchapi.io/pricing lists only hosted plans, no self-hosted option, agent-checked). data_export n/a (stateless API).]
Supadata YouTube Transcripts API
freemiumnot runnable on gaents yet
Supadata is a no-code and API-first platform for extracting transcripts, metadata, and structured content from YouTube, TikTok, Instagram, X (Twitter), and Facebook. When a user requests a YouTube transcript, Supadata first checks if YouTube provides native captions; if available, it returns them instantly with precise timestamps at a cost of 1 credit per request. If no captions exist, Supadata automatically... (Scout's description)
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
Dual-mode: fetches native captions (1 credit) with automatic AI fallback (2 credits/minute) when unavailable; no proxy configuration required (vendor-managed); supports Python/JS SDKs and REST; timestamps and speaker segmentation included; asynchronous jobs for videos >20 minutes. [2026-10-19T22:08Z cloud refresh: Dual-mode: fetches native captions (1 credit) with automatic AI fallback (2 credits/minute) when unavailable; no proxy configuration required (vendor-managed); supports Python/JS SDKs and REST; timestamps and speaker segmentation included; asynchronous jobs for videos >20 minutes. [2026-10-19 Scout T2 price/licence check: New plans since last capture: Basic ($5/month, annual), Giga ($297) and Supa ($897). Free, Pro and Mega prices unchanged. Credit rules unchanged. Plan list read from the JSON-LD on the official pricing page (dateModified 2026-09-17).]] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class and api_or_webhooks from Supadata terms (licence to call the API; software is their property). self_hostable left out (no page says hosted-only). data_export n/a (stateless API).] [2026-10-23T13:15Z cloud refresh: Fact fields pass 2 (2026-10-20): self_hostable false: hosted API only (API reference base URL https://api.supadata.ai/v1; supadata.ai/pricing lists Basic to Enterprise hosted plans and no self-hosted option, agent-checked). data_export n/a (stateless API).]
Apify YouTube Transcript Scraper
commercialnot runnable on gaents yet
Apify YouTube Transcript Scraper is a cloud-based web automation actor (serverless scraper) that extracts transcripts from YouTube videos in bulk. The tool accepts a list of YouTube video URLs or IDs, fetches native captions when available, and returns structured transcript data with timestamps and speaker segments in JSON or CSV format. For videos without captions, users can enable the AI Fallback option which... (Scout's description)
CONFIRMED 2026-10-19: Pay per event. 'Transcript scraped' $0.0007 each on Bronze to Diamond plans ($0.70 per 1,000), $0.001 each on the Apify Free plan ($1.00 per 1,000); plus 'Actor Start' $0.00005 per GB of memory per run. Price in force since 2026-07-28. The actor reads existing captions only; no AI/Whisper fallback in this actor. as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
Multiple actors on Apify marketplace; recommended: $0.5/1k option for native captions OR +Whisper AI fallback variant; no IP block issues (vendor-managed Apify infrastructure); JSON/CSV/JSONL output; ~1 min per transcript; requires Apify account. [2026-10-19T22:08Z cloud refresh: Price changed: stored $0.50 per 1,000; now 'from $0.70 / 1,000' (paid Apify plans) and $1.00 per 1,000 on the free plan. Apify's own reasonForChange: 'We have added support for scraping full channels, search results, playlists, etc'. The stored Whisper fallback ($0.012/min) is not part of this actor: its README says 'the actor extracts existing captions, not audio'. Its README still says '$0.30 per 1,000', which contradicts the live price block; the structured price block is taken as current. two_year_cost_gbp: agent-chose basis = the stored one-time job of 1,400 videos at $0.70/1k = $0.98, at £0.7502 per $1 = £0.74 (on the Apify free plan $1.40 = £1.05); excludes Apify platform plan fees.] [2026-10-19T22:08Z cloud refresh: Q4-2026 review: price rise from $0.50 to $0.70 per 1,000 on the actor page. Actor 'Last modified: a month ago'. Verdict unchanged. Pricing for this entry is carried by patches/02-price-refresh.json (fuller text, same facts); apply 02 before 03.] [2026-10-20T01:33Z cloud refresh: type proprietary -> commercial (no recurring free tier: pay per event; no recurring free tier stated; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from Apify terms; api_or_webhooks and data_export from Actor page (run via API, dataset export).]
Descript YouTube Transcription API
freemiumnot runnable on gaents yet
Descript is an end-to-end audio and video editing platform that includes state-of-the-art speech-to-text transcription powered by proprietary and open-source models. When a user uploads or links a video file, Descript transcribes the audio to text, synchronized with the video timeline, enabling line-by-line editing of the transcript to edit the video itself. The platform supports speaker identification,... (Scout's description)
CONFIRMED 2026-10-19: Free $0; Hobbyist $16/person/mo annual ($24 monthly), 10 media hours/mo; Creator $24 annual ($35 monthly), 30 media hours/mo; Business $50 annual ($65 monthly), 40 media hours/mo; Enterprise custom. UNVERIFIED: which tier includes API access. as of 2026-10-19unverified in Scout's data
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
End-to-end video-to-text platform (not YouTube-specific); STT runs locally or cloud-based; includes video editing, speaker identification, filler-word removal; requires API tier access for bulk automation; higher setup complexity than dedicated transcript APIs. [2026-10-19T23:00Z cloud refresh: 2026-10-19 price refresh: tiers confirmed; Descript now advertises 'API + MCP'. Stored 2,513-hour estimate not re-derived.] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from Descript Terms (5.1 License Grant). self_hostable and data_export left out (export help URL redirected to the help home page).]
ElevenLabs Scribe v2 (STT)
commercialnot runnable on gaents yet
ElevenLabs Scribe is an automatic speech recognition (ASR) API. Scribe v2 (GA) supports 99 languages batch; Scribe v2 Realtime supports 90+ languages at 150ms latency. Returns transcripts with speaker diarisation, word-level timestamps, and confidence scores. Integrates with ElevenLabs Conversational AI for voice-agent loops. (Scout's description)
CONFIRMED 2026-10-19 (API pricing page): Scribe v2 and Scribe v2 Medical $0.22 per hour; Scribe v2 Realtime $0.39 per hour; keyterm prompting +$0.05/h (realtime +$0.08/h); transcript editing +$0.07/h (realtime +$0.12/h). 'API usage is billed in US dollars, not credits.' 90+ languages. as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
10× more expensive than Groq Whisper Turbo ($0.40/hr vs $0.04/hr) — SKIP for Wolfram-style transcript work. SPARK isiZulu/isiXhosa land in ElevenLabs's own 'Moderate' tier (25-50% WER per vendor classification) — unusable for classroom. SPARK Afrikaans: vendor-claim 3.1% FLEURS / 5.5% Common Voice WER, claims to beat Whisper and Gemini — CONSIDER pilot 50hrs to verify. Children's voices: catalog exists but no SPARK-specific dataset eval — EVALUATE before deploying. POPIA: no SA residency; path via EU residency + Zero Retention Mode + signed Enterprise DPA — requires legal review. [2026-10-19T23:00Z cloud refresh: 2026-10-19 price refresh: batch now a flat $0.22/h on the API page (was ~$0.40 Starter to $0.28 Business); realtime is quoted per hour ($0.39), not per 1k tokens. type 'proprietary' -> 'commercial'.] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: api_or_webhooks from the speech-to-text API reference. licence_class left out (terms read contain licences from you to ElevenLabs, but I found no grant-to-you wording to quote). self_hostable left out. data_export n/a (compute API).]
FluidVoice
open-sourcenot runnable on gaents yet
Native Swift macOS dictation app doing on-device STT with an optional custom-trained local enhancement model (Fluid-1) that post-processes raw dictation — strips filler words, fixes capitalisation and formatting, normalises dates and numbers, and adapts tone to the frontmost app. Nothing leaves the machine. This is the tool in the tweet's screenshot: Pat Walls names no tool in the tweet text, and the attached... (Scout's description)
Free. Positioned explicitly as a local replacement for Wispr Flow — the $15/mo tool the tweet author was paying for. as of: no date recorded for this field
Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
GitHub stars
13058 as of: no date recorded for this field
Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
MACOS ONLY. That is disqualifying on its own — our fleet and dev hosts are Linux (kernel 6.17). The repo description states 'Windows & iOS waitlist open. Linux soon', which is roadmap language with no shipped artifact behind it, so the Linux claim is unverified and should not be planned against. Secondary flag: GPL-3.0, which is fine for using the app as-is but is a constraint if anything derived from it were ever bundled into something we distribute — worth noting since most of this category is MIT or Apache-2.0. Also note the scope: this is desktop DICTATION (voice-to-text as an input method for a human typing), not transcription-at-scale — it does not compete with Groq Whisper Turbo or AssemblyAI for corpus work, and does not compete with Vexa or Meetily for meetings.
Lock-in risk
medium
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
NOT currently in the DB. Add to Speech-to-Text / Transcription as a skip entry with the platform reason recorded, so a future session that sees FluidVoice trending does not re-litigate it. IMPORTANT: this item requires reading the tweet's attached image — the body text says only 'switched to an open source one' and names nothing. Per method, the image was downloaded from the fxtwitter media URL and read directly; the tool identity, the GPL-3.0 licence and the 'Local Wispr Flow alternative' positioning all come from that screenshot and were then independently gh-api verified. The $15/mo figure in the tweet is Wispr Flow's price, i.e. the displaced spend, not a cost of anything here. [2026-10-23T13:15Z cloud refresh: Sweep 2026-10-20: last commit 2026-10-19; licence unchanged (GPL-3.0).]
Handy
open-sourcenot runnable on gaents yet
Free, open-source, extensible push-to-talk speech-to-text that runs fully offline on macOS, Windows AND Linux. Whisper models run locally; transcribed text is injected into whatever application has focus, so it works as a system-wide dictation input method rather than a single app's feature. (Scout's description)
Raw output by design — minimal auto-punctuation, no AI clean-up layer, and a 2-5 second processing delay after you release the key. That is the honest trade against FluidVoice's Fluid-1 post-processing and against paid Wispr Flow, both of which polish the text. Accuracy is comparable to cloud services when running Whisper Large, which is also the heaviest local model, so there is a real accuracy-vs-latency-vs-VRAM dial to set per machine. Like every tool in this sub-category it is a desktop application: it serves a human at a keyboard and is not deployable on our headless agent fleet, and it does not touch batch corpus transcription.
Lock-in risk
low
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
EIGHTH ENTRY — not one of Adam's seven. Included because triaging item 5 honestly requires naming the tool that actually fits a Linux org, and burying that in another entry's notes would lose it at merge time. Add to Speech-to-Text / Transcription alongside the FluidVoice skip entry; together they give that category its first local-dictation coverage, which is currently a genuine hole (the category is otherwise six cloud APIs plus whisper.cpp plus YouTube-transcript scrapers). Cross-reference in both directions so the pair reads as one decision. Adjacent verified option if a GPL-3.0/VOSK-based alternative is ever wanted: ideasman42/nerd-dictation, 1,906 stars, GPL-3.0, last push 2025-10-10 — noticeably less maintained. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from LICENSE (MIT); self_hostable from README (local desktop app); api_or_webhooks and data_export left out (no evidence).]
Buzz
open-sourcenot runnable on gaents yet
Offline Whisper transcription GUI. Desktop app for transcribing and translating audio using OpenAI Whisper models locally. Supports recording from microphone or importing audio files. Cross-platform (macOS, Linux, Windows). Python/Qt. 20.5K stars. (Scout's description)
The most popular Whisper desktop GUI. Useful for Adam's ad-hoc transcription needs. But for pipeline/agent use, whisper.cpp or Deepgram API are better fits. Consider for desktop use only.
Limits for us
Desktop GUI app — no CLI/API for agent integration. Primarily a Whisper frontend, not a meeting intelligence tool. Overlaps with whisper.cpp (52K stars) and our existing Whisper entries. Value is in the polished desktop experience for ad-hoc transcription, not pipeline integration.
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Real-time transcription server by Collabora. Server-client architecture: server runs Whisper model, clients connect and stream audio for live transcription. WebSocket-based. Supports faster-whisper backend for GPU acceleration. Python, MIT. Linux-native. (Scout's description)
The best self-hosted real-time transcription server. MIT, Linux-native, Collabora-backed. Natural fit as the STT layer in a HoldSpeak + WhisperLive stack. Evaluate as our real-time transcription backbone.
Limits for us
Server-client architecture is a good fit for our multi-VPS setup — run the server on vpsm with GPU, stream audio from any client. Collabora (enterprise Linux company) backing adds credibility. Could serve as the STT backbone for HoldSpeak or any other capture tool.
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
[2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from LICENSE (MIT); self_hostable from README (pip/Docker); api_or_webhooks from README (OpenAI-compatible REST endpoint); data_export n/a.]
WhisperX
open-sourcenot runnable on gaents yet
Whisper with word-level timestamps and speaker diarization. Extends OpenAI Whisper with: forced alignment for word-level timestamps (via wav2vec2), voice activity detection (via pyannote.audio), and speaker diarization. Produces per-word speaker labels. BSD-2-Clause. (Scout's description)
The standard for Whisper + speaker diarization. Essential component for any meeting transcription pipeline that needs to distinguish speakers. Evaluate — this is the missing piece between 'transcribed text' and 'meeting intelligence with attribution.'
Limits for us
Speaker diarization is the key feature — knowing WHO said WHAT is critical for meeting intelligence. Complements base Whisper/whisper.cpp which only transcribe without speaker labels. pyannote.audio diarization requires a Hugging Face token for the segmentation model. 23.3K stars — massive adoption.
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Streaming wrapper around whisper.cpp. Chunk-by-chunk transcription with a growing-window strategy: each new audio slice re-transcribes the entire window rather than only the new part. (Scout's description)
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
There is no VAD at all in the runtime path. The growing-window design is O(n^2); the maintainer's own table shows latency climbing 155 -> 471 ms as one sentence grows. beam_size and best_of do not exist. Single maintainer. [2026-10-20T01:33Z cloud refresh: type None -> open-source (license text: MIT; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Sweep 2026-10-20: last commit 2026-10-09; licence unchanged (MIT).]
NVIDIA Parakeet TDT 0.6B v3
open-sourcenot runnable on gaents yet
600M-parameter TDT (token-and-duration transducer) ASR model from NVIDIA, served via the NeMo 2.4 toolkit or Transformers-from-source. 25 European languages with automatic language ID, automatic punctuation and capitalisation, and word/segment/char-level timestamps. Streaming inference supported. (Scout's description)
Free weights. Self-host cost only. ~2 GB VRAM to load, so it runs alongside an existing withheld workload on an 8 GB card rather than displacing it. Reference hardware for the published RTFx is the Open ASR Leaderboard rig, NOT a consumer GPU. as of: no date recorded for this field
Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
Evaluate, do not adopt on the leaderboard number. It dissolves the withheld-contention premise D554 accepted as a cost, but it is the only candidate with zero vocabulary control.
Limits for us
NO prompt or keyterm mechanism of any kind. There is no way to bias it toward `NestList`, `ruliad` or `hypergraph`. For Wolfram's corpus that is the single constraint that matters, and it is the reason a slower, worse-on-paper Whisper may still win. The mitigation is model-agnostic: a downstream fuzzy-map against Wolfram's ~6,000 built-in symbol names, which works on any engine's output.
Lock-in risk
low
Integration hours (estimate)
6
Feature overlap (Scout's estimate)
0.75
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
THE thing that changed in STT since 2026-05, and it changes the self-host arithmetic rather than the cloud arithmetic. 6.34% average WER on the Open ASR Leaderboard against ~7.4% for Whisper large-v3, at a quarter of the size. Three caveats carried deliberately: (1) the model card names Ampere/Blackwell/Hopper/Volta and does NOT name Turing, but T4 is on the tested list and T4 *is* Turing sm_75 — same as an RTX 2070, so expect it to work and prove it in a pilot rather than assuming; (2) 24-minute full-attention window means hour-long content needs chunking regardless; (3) 6.34% is leaderboard audio (read speech, podcasts) and does not transfer to fast math-heavy livestream. [2026-10-20T01:33Z cloud refresh: type open-weights -> open-source (CC-BY-4.0 weights (permissive open licence); rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from model card (CC-BY-4.0, attribution only; agent-chose 'permissive'). self_hostable and api_or_webhooks from model card (NeMo usage). data_export n/a (model).]
Why this link: code fact: apify-transcribe/src runs faster_whisper.WhisperModel. The listing page shows whether it can be called today.
CTranslate2 reimplementation of Whisper inference. Same weights, 4-8x faster than openai/whisper with lower VRAM, int8/float16 quantisation, VAD filtering, and (since 1.1.0) a BatchedInferencePipeline for long-audio throughput. (Scout's description)
Free. CTranslate2 backend; int8 large-v3 fits ~2.5 GB VRAM. Published unbatched throughput ~12x real-time for large-v3 on an RTX 4070; batched runs report far higher on long files. as of: no date recorded for this field
Scout, alternatives.json v3.72026-11-09Scout last_verified; no date recorded for this field itself
Already adopted in fact; the DB was just missing the row. The useful output of this entry is the version knob, not the verdict.
Limits for us
Repo last pushed 2025-11-19 (~10 months at time of writing) though releases track to v1.2.1 — a slowing cadence to watch, not a dead project. CTranslate2 itself is active (pushed 2026-08-31). Turing cards cannot use bf16; use int8_float16.
Lock-in risk
low
Integration hours (estimate)
2
Feature overlap (Scout's estimate)
0.9
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
A REAL GAP, not a new product: this is already in production in two org tools — the local STT fallback in the `yt-transcript` button and wolfram/scripts/self_host_whisper.py — and had no DB row. TUNABLES, which is the point of the entry: version pin, compute_type (int8 / int8_float16 / float16), beam_size, vad_filter, and batch_size on BatchedInferencePipeline. Wolfram's script pins ==1.0.3 (2024-07-01) while current is 1.2.1 (2025-10-31); BatchedInferencePipeline landed in 1.1.0, so the pin forfeits the batching speedup entirely and is the likeliest source of the '1.6x real-time / 5 months' estimate in D554. [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: licence_class from LICENSE (MIT); self_hostable and api_or_webhooks from README (Python library); data_export n/a.] [2026-11-09T08:48Z cloud refresh: Tweet triage 2026-11-04 #2 (2026-11-09): gh api 26,604 stars, pushed 2026-11-06; licence MIT unchanged.]
YouTube Data API v3
proprietarynot runnable on gaents yet
Official Google REST API for YouTube metadata: channel uploads playlists, video snippets (title, description, publishedAt, tags) and contentDetails (ISO-8601 duration, caption availability flag). API-key auth for public read; OAuth only for owner-scoped writes. (Scout's description)
CONFIRMED 2026-10-19: Free. Default quota: 100 search.list calls per day, 100 videos.insert calls per day, and 10,000 units per day combined for all other endpoints. search.list now has its own '100 quota per day. Each call costs 1 quota.' playlistItems.list = 1 unit; videos.list = 1 unit; captions.list = 50 units. Extra quota by request (quota and compliance audit). as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Adopt for metadata. Do not reach for it expecting transcripts.
Limits for us
No captions for content we do not own. Chapters are not a first-class field — they live as timestamp lines inside the description string returned by `snippet` and need parsing. Pair it with the `yt-transcript` button, which already handles the caption/STT half.
Lock-in risk
low
Integration hours (estimate)
3
Feature overlap (Scout's estimate)
0.4
Credited to Scout (2026-11-09, Scout last_verified). gaents does not score items.
Scout's notes
SCOPE CORRECTION worth recording: this API CANNOT fetch third-party captions. captions.list will happily return track IDs for any video, but captions.download returns 403 unless the OAuth token belongs to the channel owner — by design, not a bug. It is a metadata tool, full stop. Within that scope it is the right tool: 460 episodes cost ~20 quota units total (10 playlistItems.list + 10 videos.list) out of 10,000/day. The 'brittle to quota' reputation comes entirely from search.list at 100 units — enumerate the uploads playlist instead and quota is a non-issue with ~500x headroom. [2026-10-19T22:08Z cloud refresh: 2026-10-19 freshness check: search.list quota model changed. Stored text said search.list = 100 units out of the 10,000/day pool. The official quota page now gives search.list a separate allowance of 100 calls per day, 1 quota each; the 10,000 units/day pool covers 'all other endpoints'. Other unit costs unchanged.] [2026-10-23T13:15Z cloud refresh: type left OPEN (free API; extra quota is by request form, not a paid tier). See out_types.md.] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: Correction: stored type 'proprietary' is not an allowed type value; set to 'commercial' (closed vendor service, free within quota). licence_class from YouTube API Services Terms; self_hostable false: hosted Google API, no self-hosted edition (getting-started page); api_or_webhooks from the same page; data_export n/a. [merge: `type` not set here. Patch 08 left this row OPEN (a free closed API fits none of the four type values).] [merge: `api_or_webhooks` is carried by patches/08-open-types.json.]]
youtube-transcript-api (jdepoix)
open-sourcenot runnable on gaents yet
Python library that fetches YouTube's own caption tracks (manual and auto-generated) without an API key and without the captions.download ownership restriction, by hitting the same endpoints the player uses. Returns timestamped segments; supports translation tracks and a pluggable proxy config. (Scout's description)
Free. Cost is proxy spend if YouTube IP-blocks the caller — the library itself takes a proxy config rather than bundling one. as of: no date recorded for this field
Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
GitHub stars
8611 as of: no date recorded for this field
Scout, alternatives.json v3.72026-10-23T13:15ZScout patch_history.at; no date recorded for this field itself
Consider, do not adopt. The button is the org path; this is the in-process fallback.
Limits for us
DUPLICATES A LIVE ORG TOOL. `yt-transcript` already does captions-first fetching plus PO-token/cookie handling plus a per-video cache plus Groq STT fallback; this library does the caption fetch alone and has no bot-block story beyond 'bring your own proxy'. Flagged rather than recommended: reach for it only if something needs captions in-process without shelling out to the button.
Lock-in risk
low
Integration hours (estimate)
1
Feature overlap (Scout's estimate)
0.55
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
The clean library for the caption layer when you need it OUTSIDE the `yt-transcript` button. Actively maintained (pushed 2026-09-10) with only 32 open issues on 8.3k stars, which is a healthier ratio than most scraping libraries. TUNABLES: proxy config, language preference list, preserve_formatting, and manual-vs-generated track preference. [2026-10-23T13:15Z cloud refresh: Sweep 2026-10-20: last commit 2026-09-09; licence unchanged (MIT).]
Mistral Voxtral Mini Transcribe V2
commercialnot runnable on gaents yet
Mistral's hosted transcription endpoint, also reachable through withheld. Batch and realtime tiers; positioned as a cheaper European-hosted alternative to OpenAI's audio endpoints. (Scout's description)
CONFIRMED 2026-10-19: Voxtral Mini Transcribe V2 API $0.003/min ($0.18/hour); Voxtral Realtime API $0.006/min, also open weights under Apache 2.0. At Wolfram's 2,513-hour corpus: $452.34 batch (agent arithmetic: 150,780 min x $0.003). as of 2026-10-19
Scout, alternatives.json v3.72026-10-19date written in Scout's pricing text
Watch. Real product, wrong axis for the corpus that prompted the look.
Limits for us
No verified independent WER benchmark on the Open ASR Leaderboard at time of writing — the accuracy claims are vendor-side. Nothing here beats the Groq/AssemblyAI pair on either the cost axis or the vocabulary axis, so there is no current reason to test it.
Lock-in risk
medium
Integration hours (estimate)
2
Feature overlap (Scout's estimate)
0.7
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Priced at exactly OpenAI's gpt-4o-mini-transcribe tier and 4.5x Groq whisper-large-v3-turbo ($0.18/h vs $0.04/h). For the 2,513-hour corpus that is $452 against $101 for the same job. No published keyterm or domain-vocabulary mechanism found at time of writing, which is the axis Wolfram's corpus is actually constrained on. [2026-10-19T23:00Z cloud refresh: 2026-10-19: type 'proprietary' is not an allowed value; set to commercial (Mini Transcribe V2 is API-only; Realtime sibling is Apache-2.0 open weights).] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: launch post names only Voxtral Realtime as open weights (Apache 2.0); Mini Transcribe V2 is sold as a transcription API ($0.003/min), so licence_class proprietary (agent reading: no open licence stated for Mini Transcribe V2). self_hostable true: the post says both models support on-premise setups (enterprise). api_or_webhooks true: offered as a transcription API. data_export n/a (pure compute API).]
insanely-fast-whisper
open-sourcenot runnable on gaents yet
CLI wrapper over HF Transformers Whisper with Flash Attention 2 and chunked batching, marketed on transcribing 150 minutes of audio in under 98 seconds on an A100. (Scout's description)
Skip on maintenance and on hardware, not on merit.
Limits for us
Unmaintained for ~11 months, and architecturally excluded from the one piece of hardware D554 is about.
Lock-in risk
low
Integration hours (estimate)
2
Feature overlap (Scout's estimate)
0.6
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Recorded as a skip WITH EVIDENCE so a future session does not re-run this search (the FluidVoice pattern). Last push 2025-10-25 — roughly 11 months stale at 2026-09-14, with 116 open issues. It was a legitimate speed record in 2024; faster-whisper's BatchedInferencePipeline and Parakeet TDT have since taken that ground from a maintained codebase. Flash Attention 2 also requires Ampere or newer, so it cannot run on the Turing RTX 2070 that prompted the question. [2026-10-23T13:15Z cloud refresh: Sweep 2026-10-20: last commit 2024-05-27; licence unchanged (Apache-2.0).]
Distil-Whisper
open-sourcenot runnable on gaents yet
Hugging Face's knowledge-distilled Whisper variants — ~6x faster and ~49% smaller than large-v2, English-only, with a claimed WER within 1% on out-of-distribution audio. (Scout's description)
Skip. Superseded by Parakeet TDT 0.6B v3 on every axis that mattered.
Limits for us
Effectively unmaintained. English-only. Superseded on both the speed and accuracy axes by a model a quarter of its size.
Lock-in risk
low
Integration hours (estimate)
3
Feature overlap (Scout's estimate)
0.6
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Recorded as a skip WITH EVIDENCE so it is not re-litigated. Last push 2025-01-08 — about 20 months stale at 2026-09-14, with 75 open issues. It is a research artefact that shipped and stopped. Parakeet TDT 0.6B v3 now occupies the same niche (small, fast, English-strong) from a maintained source and with a better leaderboard number, so the distillation trade is no longer worth making. [2026-10-23T13:15Z cloud refresh: Sweep 2026-10-20: last commit 2025-01-08; licence unchanged (MIT).]
Workers AI Whisper-large-v3-turbo
open-sourcenot runnable on gaents yet
See research/cloudflare-offerings-deep-review-2026-10-02.md (section for Workers AI Whisper-large-v3-turbo). (Scout's description)
One triage pass; re-verify stars, licence and price before acting.
Limits for us
See research/cloudflare-offerings-deep-review-2026-10-02.md (section for Workers AI Whisper-large-v3-turbo).
Credited to Scout (2026-10-23T13:15Z, Scout patch_history.at). gaents does not score items.
Scout's notes
Merged from research/cloudflare-offerings-deep-review-2026-10-02.md by merge_triage_table.py. [2026-10-20T01:33Z cloud refresh: type None -> open-source (license text: MIT; rule set agent-chose 2026-10-20, Adam to rule)] [2026-10-23T13:15Z cloud refresh: Fact fields 2026-10-20: api_or_webhooks from the model page; licence_class left out (row mixes MIT weights and the Cloudflare service; model page shows no licence text); self_hostable left out (row is the hosted service); data_export n/a.]
OpenAI gpt-transcribe
commercialnot runnable on gaents yet
OpenAI's speech-to-text model for recorded audio, released 2026-07-28 and named by OpenAI as the replacement for whisper-1, gpt-4o-transcribe and gpt-4o-mini-transcribe (all shut down 2027-02-26). Same /v1/audio/transcriptions endpoint; also final transcripts of committed turns in Realtime transcription sessions. Accepts a free-text prompt, keyword hints and multiple language hints for domain terms and... (Scout's description)
Cloud research pass 2026-10-19; re-verify price and licence before acting.
Limits for us
Hosted only; audio leaves our servers. Files up to 25 MB per request. OpenAI says to use a specialised model for speaker labels, word timestamps, subtitle formats or translation; its diarisation model (gpt-4o-transcribe-diarize) is itself being removed on 2027-02-26. No Batch API. At $0.27/hour it costs about 6.75x Groq whisper-large-v3-turbo ($0.04/hour).
Lock-in risk
low
Integration hours (estimate)
2
Feature overlap (Scout's estimate)
0.9
Credited to Scout (2026-10-19, Scout evidence.read_at). gaents does not score items.
https://developers.openai.com/api/docs/guides/speech-to-text (read 2026-10-19): “Start with gpt-transcribe . This is the recommended model for transcribing recorded speech in its original language. Use a specialized model only if you need speaker labels, word timestamps, subtitle formats, or translation into English. Files can be up to 25 MB.”
Scout's notes
Model ID gpt-transcribe (only snapshot listed). Released per the OpenAI changelog, Jul 28 2026. two_year_cost_gbp: agent-chose basis = one pass over the 2,513-hour corpus named in the Voxtral entry: 2,513 h x 60 x $0.0045 = $678.51, GBP at £0.7502 per $1 (brief rate, 2026-10-19) = £509.02. Not a recurring cost estimate. integration_hours 2 (drop-in model name on the endpoint we already call for whisper-1), lock_in_risk low (OpenAI-compatible transcription endpoint, many alternatives in this category), feature_overlap 0.9 (against OpenAI Whisper API / Groq Whisper) and verdict evaluate are agent judgements. Why evaluate: keyword hints address the vocabulary axis noted on the Voxtral entry; worth a WER test against Groq turbo on that corpus.
AuroraASR-1M
open-sourcenot runnable on gaents yet
949K-parameter Conformer-CTC English ASR model (FP32 3.81 MB, INT8 1.06 MB), trained only on LibriSpeech 960h. Runs through Transformers (trust_remote_code). Card reports 436x real time on one CPU thread, 1362x on an RTX PRO 6000 Blackwell. (Scout's description)
Cloud research pass 2026-10-19; re-verify price and licence before acting.
Limits for us
LibriSpeech WER 11.43% test-clean / 26.50% test-other (Whisper tiny.en ~5.6% clean); English only; no punctuation or casing; trained on read audiobooks only, so call/meeting audio will be worse. Needs trust_remote_code. Not a replacement for faster-whisper/Parakeet/Deepgram on our transcripts; only fits wake-word/command or on-device triage.
Lock-in risk
low
Integration hours (estimate)
1
Feature overlap (Scout's estimate)
0.2
Credited to Scout (2026-10-23, Scout evidence.read_at). gaents does not score items.
Tweet triage 2026-10-23 #7 (TeksEdge). Card numbers match the tweet. Runs on any of our boxes (main CPU, withheld, air1/air2) - hardware is not the constraint, accuracy is.
Moonshine (moonshine-ai)
open-sourcenot runnable on gaents yet
Low-latency on-device streaming STT (plus intent recognition and TTS) for voice agents. Ships models from tiny micro builds (~1.3 MiB command-recognition CNN) up to streaming models it says beat Whisper Large V3. (Scout's description)
legacy non-English models under the Moonshine Community Licence
Scout, gaents-proven-2026-11-18.md A22026-11-18Scout review date
Scout's judgement
Verdict
evaluate
Reasoning (de-assume)
Cloud research pass 2026-10-19; re-verify price and licence before acting.
Limits for us
README gives no WER table; accuracy claim points to the HF Open ASR leaderboard (not checked here). Micro models are command recognition, not open transcription. Benchmark on our audio before any use.
Lock-in risk
low
Integration hours (estimate)
2
Feature overlap (Scout's estimate)
0.5
Credited to Scout (2026-10-23, Scout evidence.read_at). gaents does not score items.
Tweet triage 2026-10-23 #7: the stronger tiny/edge alternative to AuroraASR-1M. Fathom to A/B against faster-whisper on our call samples.
Vosk
open-sourcenot runnable on gaents yet
Offline Kaldi-based speech recognition API for Android, iOS, Raspberry Pi and servers (Python, Java, C#, Node); small 40-50 MB models in 20+ languages. (Scout's description)
Cloud research pass 2026-10-19; re-verify price and licence before acting.
Limits for us
Older architecture; accuracy below Whisper-family models on general audio. Useful only where a multilingual offline CPU model is required.
Lock-in risk
low
Integration hours (estimate)
2
Feature overlap (Scout's estimate)
0.4
Credited to Scout (2026-10-23, Scout evidence.read_at). gaents does not score items.
Scout's evidence (1 quotes)
https://github.com/alphacep/vosk-api (read 2026-10-23): “Offline speech recognition API for Android, iOS, Raspberry Pi and servers with Python, Java, C# and Node”
Scout's notes
Tweet triage 2026-10-23 #7 alternative set (offline CPU STT).