# Turn a long recording into plain text

> Turn a long recording into plain text: 5 companies fully do it, 20 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.long-recording-batch/
- Markdown: https://www.anchorterminal.com/niches/speech-to-text.long-recording-batch/index.md (~2,200 tokens)
- Slim: https://www.anchorterminal.com/niches/speech-to-text.long-recording-batch/index.min.md (~580 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/speech-to-text.long-recording-batch/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 |
| --- | --- | --- | --- | --- |
| 5 | 20 | 42 | 14 | 7 |

## What counts as done

The whole recording is transcribed in one job or an automatic chunking the tool manages.

Measured, for our own tests: full text to the last spoken word, with the final word's timestamp within 5 seconds of the audio's duration, and no repeated or invented sentences over silent stretches.

How agents ask for it: transcribe this two-hour podcast; get me the text of this lecture video; transcribe these 40 council meeting recordings

## 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 accepts files under 5 hours and 500 MB synchronously; batch takes 1 GB files from a URL with word timestamps.
- **Deepgram Speech-to-Text (Nova-3, Flux) (Deepgram, Inc.)**: Pre-recorded accepts files to 2 GB by upload or URL with a callback for long jobs; the duration field is returned on every response.
- **ElevenLabs Scribe Speech to Text API (Eleven Labs Inc.)**: Files to 5 GB and 15 hours per request with audio_duration_secs returned; webhook delivery for long jobs.
- **IBM Watson Speech to Text (International Business Machines Corporation)**: /v1/recognitions takes up to 1 GB per request with callbacks; results are kept one week by default.
- **Rev AI Speech-to-Text API (Rev.com, Inc.)**: Jobs take media to 17 hours (6 for Telugu) and 2 GB from a URL, with webhooks.

## 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 take files to 28,800 seconds and 2 GB, but input must sit in S3 and acc-06 reports timestamp drift on long files.
- **AssemblyAI Speech-to-Text (Universal) (AssemblyAI, Inc.)**: One job takes files to 10 hours and 5 GB, but acc-30 shows a one-hour file losing half its text on the legacy Best tier; the maintainer advised universal-3-5-pro and checking the last word's end time.
- **Cartesia Ink (Cartesia AI, Inc.)**: POST /stt transcribes ink-whisper files with a duration field, but no file size or duration limit is documented and ink-2 has no batch endpoint.
- **Cohere Transcribe 03-2026 (Cohere Inc.)**: The Transformers processor chunks long audio and reassembles by audio_chunk_index, but Cohere's API caps files at 25 MB and the blog says the model transcribes non-speech sounds without a VAD.
- **Giggy Speech to Text (Giggy)**: Free Batch stops at 25 minutes; paid Fast takes recordings to 12 hours but only as 16 kHz mono PCM WAV, and the transcript is fetched from a signed URL that can disappear after retention cleanup.
- **Gladia Speech-to-Text API + MCP (Gladia SAS)**: Pre-recorded files are capped at 135 minutes and 1,000 MB with an error beyond that, so Gladia advises splitting into about 60-minute chunks yourself.
- **Google Cloud Speech-to-Text (Google LLC)**: BatchRecognize takes files to 8 hours but only from Cloud Storage; acc-05 and acc-06 report Chirp repeating filler text on silence and skipping speech.
- **Groq Speech-to-Text (Groq LLC)**: Uploads stop at 25 MB free or 100 MB on the Developer plan; url avoids the attachment cap but the plan size limit still forces client-side chunking of long recordings.
- **Inworld STT (Theai, Inc. dba Inworld AI)**: Async jobs take 512 MB of audio or video, but results and the recording are deleted after 7 days, so the agent must collect in time.
- **Ming-3.0-asr-tiny (inclusionAI, Ant Group (named in the site footer of inclusion-ai.org; the LICENSE copyright line reads inclusionAI))**: Long audio is encoded in chunks with --audio-chunk-frames, within an 8,192-token context shared by audio and text, so hours of audio need the caller to split.
- **Ming-3.0-audio-flash (inclusionAI, Ant Group (named in the inclusion-ai.org footer; no registered company named))**: The ASR task runs through vLLM-Omni or SGLang on four to eight 80 GB GPUs; the card warns long audio may OOM the four-GPU profile and documents no chunking.
- **Ming-3.0-audio-tiny (inclusionAI (Ant Group))**: ASR runs on one GPU through SGLang-Omni or vLLM-Omni with inclusionAI's code; no chunking or duration limit is documented for long files.
- **Mistral Voxtral Transcribe (Mistral AI (RCS Paris 952 418 325))**: Batch takes file, file_url or file_id; the up-to-3-hours figure is only on the announcement page, not in the API reference read.
- **Nemotron 3.5 ASR Streaming 0.6B (NVIDIA Corporation)**: Offline transcription is documented in NeMo and Transformers with chunked streaming, but no duration limit or long-file test is published.
- **OpenAI Speech to Text (OpenAI OpCo, LLC)**: Files stop at 25 MB and the caller splits; acc-19 shows gpt-4o-transcribe truncating at output_tokens 2048 with no error, unchecked for gpt-transcribe.
- **Parakeet TDT 0.6B v3 (NVIDIA Corporation)**: Audio past about 24 minutes needs NeMo switched to local attention, and a discussion reports CUDA out of memory on long audio; no managed chunking in Transformers.
- **Qwen3-ASR-0.6B (Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check))**: The report covers single recordings to 20 minutes and max_new_tokens must be raised for long audio, or output is cut; longer files need caller chunking.
- **Qwen3-ASR-1.7B (Alibaba Cloud (the registrant of qwen.ai per RDAP, carried from the Qwen3.8-27B check))**: The report covers single recordings to 20 minutes and max_new_tokens must be raised for long audio, or output is cut; longer files need caller chunking.
- **Smallest.ai Pulse STT (Smallest, Inc.)**: Pre-recorded accepts files with webhook delivery; Pulse Pro caps audio at 250 MB or 50 MB with a webhook and returns 413 beyond, and no duration cap is stated.
- **Soniox Speech-to-Text (Soniox Inc.)**: Async files are capped at a fixed 300 minutes and must be deleted by the caller; files and transcripts expire after 30 days.
- **Speechmatics Speech-to-Text (Cantab Research Ltd)**: Files to 1 GB in the request body, larger by URL, but transcripts expire after 7 days and no duration cap is stated.
- **Underdog Bark 3.0 (Conway Research)**: The card says split recordings into a few minutes on 8 or 16 GB machines and that long-recording segmentation has not been qualified; a 10-minute file peaked near 15 GB.
- **Voxtral Mini 4B Realtime 2602 (Mistral AI (RCS Paris 952 418 325))**: vLLM's default max model length covers about 3 hours, but acc-22 reports output cut short unless max_new_tokens is at least duration over 80 ms, and acc-27 shows the Docker image failing to load it.

- not this job: MiniMax Speech-to-Text (Nanonoble Pte. Ltd.): Audio over 500 seconds or 50 MB returns 400, so a long recording cannot run as one job and the tool manages no chunking.

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

## What stops agents finishing

- silent truncation at an output token cap (acc-19: gpt-4o-transcribe stopped at about 8 to 9 minutes with no error)
- hard duration caps far below the recording (MiniMax 500 seconds, Google sync 1 minute, Groq 25 MB free)
- input must sit in the vendor's own storage (Amazon S3, Google Cloud Storage)
- transcript URL expires before the agent collects it (Speechmatics 7 days, Inworld 7 days, Giggy signed URL)

## Around it

Before: media.download-video, storage.fetch-file. After: llm.correct-proper-nouns, llm.summarise, search.index-text.

## Where this comes from

Weight 14 of 100 in Speech-to-text: 7 of 32 accounts (acc-09, acc-13, acc-15, acc-16, acc-19, acc-30, acc-32); truncation and silence hallucination were the two failures named

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