# Dictate a short clip into text for an agent or editor

> Dictate a short clip into text for an agent or editor: 8 companies fully do it, 17 with a limit, 42 carry the tag. Every listing in Speech-to-text with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/speech-to-text.dictation-to-agent/
- Markdown: https://www.anchorterminal.com/niches/speech-to-text.dictation-to-agent/index.md (~2,100 tokens)
- Slim: https://www.anchorterminal.com/niches/speech-to-text.dictation-to-agent/index.min.md (~580 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/speech-to-text.dictation-to-agent/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

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 8 | 17 | 42 | 10 | 4 |

## What counts as done

Clean punctuated text back for a 30-second clip.

Measured, for our own tests: the text arrives within 2 seconds of the clip ending, with no filler text or hallucinated sentence on a clip that is partly silent.

How agents ask for it: transcribe this 20-second voice note; I'll speak my prompt, type it for me; turn this voice message into text

## Does the job

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

- **Azure AI Speech speech-to-text (Microsoft Corporation)**: Fast transcription returns the text synchronously for short files; real-time recognize-once is the SDK path. No silence fixture was run.
- **Deepgram Speech-to-Text (Nova-3, Flux) (Deepgram, Inc.)**: Pre-recorded is synchronous over HTTP and returns the transcript in the response; Pipecat puts Nova-3 finals at 247 ms. No silence fixture was run.
- **Inworld STT (Theai, Inc. dba Inworld AI)**: Synchronous POST transcribe returns the transcript for requests to 16 MiB; no silence fixture was run.
- **MiniMax Speech-to-Text (Nanonoble Pte. Ltd.)**: Synchronous transcription of clips to 500 seconds with a language header; the guide claims tuning against hallucination, which is a vendor claim.
- **Mistral Voxtral Transcribe (Mistral AI (RCS Paris 952 418 325))**: Synchronous /v1/audio/transcriptions returns the text; acc-09 reports no hallucination on silence for daily use.
- **Nemotron 3.5 ASR Streaming 0.6B (NVIDIA Corporation)**: A 0.6B model running faster than real time on a small GPU or laptop via Transformers pipeline or NeMo-Speech.cpp; silence behaviour is undocumented.
- **OpenAI Speech to Text (OpenAI OpCo, LLC)**: Synchronous /v1/audio/transcriptions with gpt-transcribe returns the text; no silence fixture was run and acc-15 used an LLM pass to fix Whisper hallucinations.
- **Parakeet TDT 0.6B v3 (NVIDIA Corporation)**: acc-17 reports near-instant local dictation on a MacBook; Transformers 5.10 pipeline runs it. Silence behaviour undocumented.
- **Smallest.ai Pulse STT (Smallest, Inc.)**: POST /waves/v1/stt/ is synchronous without webhook_url and returns the transcript; silence behaviour untested.

## 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.

- **Amazon Transcribe (Amazon Web Services, Inc.)**: Batch jobs are asynchronous with S3 input and polling, so a 30-second clip is not back within 2 seconds; streaming works but the median final latency is 1,136 ms.
- **AssemblyAI Speech-to-Text (Universal) (AssemblyAI, Inc.)**: Pre-recorded is an asynchronous job with polling and a 403 rate limit; the streaming path returns finals in about 300 ms but needs a WebSocket. Silence behaviour untested.
- **Cartesia Ink (Cartesia AI, Inc.)**: Batch ink-whisper returns a transcript synchronously; realtime finals arrive in about 300 ms. Silence behaviour untested, and realtime bills silence.
- **Cohere Transcribe 03-2026 (Cohere Inc.)**: A 2B model transcribes a clip quickly on a GPU, but the blog says it is eager to transcribe non-speech sounds, so a partly silent clip needs a VAD in front.
- **ElevenLabs Scribe Speech to Text API (Eleven Labs Inc.)**: POST /v1/speech-to-text is synchronous and takes clips from 100 ms, but a request under one credit bills one credit and no silence fixture was run.
- **Giggy Speech to Text (Giggy)**: Recordings from 0.5 seconds are accepted, but every recorded request is an HTTP 202 then polling, and input must be converted to 16 kHz mono WAV first.
- **Gladia Speech-to-Text API + MCP (Gladia SAS)**: Pre-recorded is an asynchronous job (init, upload, poll or callback), so a 30-second clip is not back within 2 seconds; live sessions suit it better.
- **Google Cloud Speech-to-Text (Google LLC)**: Synchronous Recognize takes up to 1 minute or 10 MB and returns directly, but acc-05 saw repeated filler text on noisy or silent audio and a 403 on 20 seconds of silence.
- **Groq Speech-to-Text (Groq LLC)**: Synchronous transcription of a short file is fast on Groq, but every request bills at least 10 seconds and verbose_json exposes no_speech_prob rather than suppressing silence.
- **IBM Watson Speech to Text (International Business Machines Corporation)**: Synchronous /v1/recognize returns text for a short clip, but the 30-second inactivity timeout and no silence fixture leave the silence case unchecked.
- **Ming-3.0-asr-tiny (inclusionAI, Ant Group (named in the site footer of inclusion-ai.org; the LICENSE copyright line reads inclusionAI))**: A short clip transcribes in one pass on a CUDA GPU, but the model needs 17.7 GB of BF16 weights and no silence behaviour is documented.
- **Ming-3.0-audio-tiny (inclusionAI (Ant Group))**: A clip transcribes on one GPU with about 25 GB free, but loading runs repository code and silence behaviour is undocumented.
- **Qwen3-ASR-0.6B (Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check))**: Fast on a small GPU through vLLM, but acc-25 reports noise returned as filler interjections and heavy repetition on non-speech sound.
- **Qwen3-ASR-1.7B (Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check))**: Fast on a small GPU through vLLM, but acc-25 reports noise returned as filler interjections and heavy repetition on non-speech sound.
- **Rev AI Speech-to-Text API (Rev.com, Inc.)**: Async jobs poll or call back, so a 30-second clip is not back within 2 seconds, and every job bills at least 15 seconds.
- **Soniox Speech-to-Text (Soniox Inc.)**: The async API is a job with upload and polling; the realtime socket returns finals in about 266 ms but needs a stream. Silence behaviour untested.
- **Speechmatics Speech-to-Text (Cantab Research Ltd)**: Batch is a job with polling (5-second default wait) or notifications, so a clip is not back within 2 seconds; realtime finals arrive in 369 to 495 ms.
- **Underdog Bark 3.0 (Conway Research)**: About 1.5 seconds per minute of audio on an M2 per the card, but every route runs Conway's Python and the first run needs the network and about 6 GB to extract.
- **Voxtral Mini 4B Realtime 2602 (Mistral AI (RCS Paris 952 418 325))**: Runs on a 16 GB GPU in vLLM or as a Q4 GGUF in llama.cpp, but output is plain text and silence behaviour is undocumented.

- not this job: Ming-3.0-audio-flash (inclusionAI, Ant Group (named in the inclusion-ai.org footer; no registered company named)): Needs a multi-GPU server (259 GB BF16), so not a path for a waiting user on a laptop or one GPU.

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

## What stops agents finishing

- every request billed as at least 10 seconds (Groq), so many short clips cost more than their length
- tool must have 16 kHz mono WAV and nothing else (Giggy)
- noise or breath transcribed as filler words (acc-25)

## Around it

Before: voice.capture-microphone. After: llm.prompt, editor.insert-text.

## Where this comes from

Weight 10 of 100 in Speech-to-text: 4 of 32 accounts (acc-05, acc-17, acc-18, acc-25); heavily represented in Hacker News dictation-app threads, where local models dominate

As data: https://www.anchorterminal.com/api/v1/use-cases/speech-to-text.dictation-to-agent.json, or the MCP tool get_use_case with id speech-to-text.dictation-to-agent. 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
