# Transcribe on the user's own hardware without sending audio out

> Transcribe on the user's own hardware without sending audio out: 2 companies fully do it, 4 with a limit, 120 carry the tag. Every listing in Speech-to-text with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/speech-to-text.self-hosted-local/
- Markdown: https://www.anchorterminal.com/niches/speech-to-text.self-hosted-local/index.md (~1,800 tokens)
- Slim: https://www.anchorterminal.com/niches/speech-to-text.self-hosted-local/index.min.md (~330 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/speech-to-text.self-hosted-local/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Two or three companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 2 | 4 | 120 | 10 | 8 |

## What counts as done

The model loads from a released runtime (Transformers, vLLM, llama.cpp, MLX or similar) without trust_remote_code.

Measured, for our own tests: it transcribes a 10-minute file faster than real time on the stated hardware, and matches the published WER within 2 points on a public set.

How agents ask for it: transcribe this locally, nothing leaves the machine; run a speech model on my Mac or one GPU; cheap transcription for hundreds of hours

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Parakeet TDT 0.6B v3 (NVIDIA Corporation)**: Transformers pipeline loads the safetensors without remote code; the Open ASR Leaderboard row (6.32 mean WER in the repository's eval file, 4.94 on the cleaned sets per a third party) is independent of NVIDIA.
- **Qwen3-ASR-0.6B (Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check))**: vLLM 0.32 and Transformers 5.13 (-hf repository) load it without remote code; a third-party card quotes its leaderboard row at 5.16 average on the cleaned sets, but the publisher figure was not compared.
- **Qwen3-ASR-1.7B (Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check))**: vLLM 0.32 and Transformers 5.13 (-hf repository) load it without remote code; the Open ASR Leaderboard README lists Qwen3-ASR-1.7B as an evaluated model but its row could not be read.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Cohere Transcribe 03-2026 (Cohere Inc.)**: CohereAsrForConditionalGeneration loads in Transformers without remote code, but the Hugging Face repository is gated behind a human-accepted licence, and the vLLM recipe uses --trust-remote-code.
- **Ming-3.0-asr-tiny (inclusionAI, Ant Group (named in the site footer of inclusion-ai.org; the LICENSE copyright line reads inclusionAI))**: Loads only with trust_remote_code in Transformers or SGLang-Omni main; the published WER is the publisher's own with no leaderboard row.
- **Nemotron 3.5 ASR Streaming 0.6B (NVIDIA Corporation)**: Transformers 5.13 pipeline loads the safetensors without remote code; the GGUF needs NVIDIA's NeMo-Speech.cpp, and no independent WER reproduction was found.
- **Underdog Bark 3.0 (Conway Research)**: No runtime reads model.bark without Conway's loader, and the card itself says the Bark rows are not on the Open ASR Leaderboard; accuracy rests on Conway's own runs.
- **Voxtral Mini 4B Realtime 2602 (Mistral AI (RCS Paris 952 418 325))**: vLLM and Transformers 5.2 load it without remote code, but acc-22 could not reproduce the paper's WER (FLEURS EN 13.00 against 4.90) and acc-27 hit an outdated Transformers in the vLLM image.

- not this job: Amazon Transcribe (Amazon Web Services, Inc.): A hosted AWS service with no weights; audio must go to AWS.
- not this job: AssemblyAI Speech-to-Text (Universal) (AssemblyAI, Inc.): Hosted API only, no weights.
- not this job: Azure AI Speech speech-to-text (Microsoft Corporation): Hosted Azure service; containers exist but are not open weights and still bill through Azure.
- not this job: Cartesia Ink (Cartesia AI, Inc.): Hosted API only.
- not this job: Deepgram Speech-to-Text (Nova-3, Flux) (Deepgram, Inc.): Self-hosted Deepgram is a licensed container, not open weights.
- not this job: ElevenLabs Scribe Speech to Text API (Eleven Labs Inc.): Hosted API only.
- not this job: Giggy Speech to Text (Giggy): Hosted API only.
- not this job: Gladia Speech-to-Text API + MCP (Gladia SAS): Hosted API only.
- not this job: Google Cloud Speech-to-Text (Google LLC): Hosted Google Cloud service.
- not this job: Groq Speech-to-Text (Groq LLC): Hosted inference of OpenAI's Whisper; no weights from Groq.
- not this job: IBM Watson Speech to Text (International Business Machines Corporation): IBM Cloud service; the on-premises Cloud Pak is licensed software, not open weights.
- not this job: Inworld STT (Theai, Inc. dba Inworld AI): Hosted API only.
- not this job: Ming-3.0-audio-flash (inclusionAI, Ant Group (named in the inclusion-ai.org footer; no registered company named)): Every loading path uses trust_remote_code or inclusionAI's forks, needs a multi-GPU server, and no independent WER exists.
- not this job: Ming-3.0-audio-tiny (inclusionAI (Ant Group)): Needs trust_remote_code or SGLang-Omni main, no released runtime, and only the publisher's WER.
- not this job: MiniMax Speech-to-Text (Nanonoble Pte. Ltd.): Hosted API only.
- not this job: Mistral Voxtral Transcribe (Mistral AI (RCS Paris 952 418 325)): This listing is the hosted API; the open weights are listed separately as voxtral-mini-4b-realtime-2602.
- not this job: OpenAI Speech to Text (OpenAI OpCo, LLC): Hosted API; OpenAI's open Whisper weights are not this listing.
- not this job: Rev AI Speech-to-Text API (Rev.com, Inc.): Hosted API only.
- not this job: Smallest.ai Pulse STT (Smallest, Inc.): Hosted API only.
- not this job: Soniox Speech-to-Text (Soniox Inc.): Hosted API only.
- not this job: Speechmatics Speech-to-Text (Cantab Research Ltd): Containers are licensed deployments, not open weights.

3 listings in the category were added after this reading and aren't marked yet.

## What stops agents finishing

- no released runtime loads it (Ming, Nemotron GGUF needs NeMo-Speech.cpp, Underdog Bark needs the publisher's loader)
- gated download needing a human to accept terms (Cohere Transcribe)
- serving stack ships an old Transformers and the model will not start (acc-27)
- generation cap set too low silently truncates (acc-22: max_new_tokens must be at least duration / 80 ms)
- CUDA out of memory on long audio (Parakeet discussion #41)

## Around it

Before: model.download-weights, audio.resample-16k. After: diarisation.pyannote, llm.summarise.

## Where this comes from

Weight 10 of 100 in Speech-to-text: 8 of 32 accounts ran open weights (acc-08, acc-12, acc-13, acc-17, acc-22, acc-25, acc-27, acc-32); counted here where the local run itself was the point

As data: https://www.anchorterminal.com/api/v1/use-cases/speech-to-text.self-hosted-local.json, or the MCP tool get_use_case with id speech-to-text.self-hosted-local. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
