# Transcription & Speech-to-Text API

**Audo AI**'s speech-to-text and video transcription API (`transcribe_audio`) turns audio and video recordings up to **10 GB** and **10 hours** into structured text with millisecond word-level timestamps (`start_ms`, `end_ms`) and automatic speaker diarization (up to **8 speakers** on by default)—bypassing OpenAI Whisper's 25 MB upload wall with zero client-side chunking or `ffmpeg` audio extraction. In production, a 2-minute file finishes in **under 10 seconds** and recent 1-hour files complete in **under 2 minutes**. Call it with one `curl` command and no API key via `POST https://audo.ai/api/try/transcribe`, from backend jobs via the [REST API](https://audo.ai/docs/api?utm_source=markdown&utm_medium=agent#rest-api) (`/api/v1/jobs`), or from Claude Code, Cursor, Codex, and ChatGPT via the hosted [MCP server](https://audo.ai/docs/mcp?utm_source=markdown&utm_medium=agent) (`https://audo.ai/mcp`).

- **Price & keyless free tier:** **$0.03/min ($1.80/hr)** pay-as-you-go from $5 (**$1.13–$1.44/hr** with bonus credits on $10–$100 packs), no monthly subscription, credits never expire, no multichannel surcharge, and failed jobs are never charged. **First minute** of each file is free **3x/day** with zero API key via `/api/try/transcribe` (see [pricing](https://audo.ai/pricing?utm_source=markdown&utm_medium=agent)).
- **Measured speed:** A 2-minute file returns in **under 10 seconds**; recent 1-hour recordings finish in **under 2 minutes** on warm GPUs.
- **Audio & video inputs (up to 10 GB):** Audio (WAV, MP3, M4A, AAC, FLAC, OGG, Opus, WebM, AIFF) and raw video (MP4, MOV, MKV, WebM, AVI) up to **10 GB** and **10 hours** per file (single-request upload up to 100 MB, or multipart upload over 16 MB).
- **Languages & speaker diarization:** **60 languages** detected automatically or set with ISO code `language` (such as `en`, `es`, or `ja`), with speaker labels (`Speaker 1` through `Speaker 8`) enabled by default.
- **Outputs & transcript reuse:** Word-timed **JSON**, speaker-turn **TXT**, **SRT**, and **VTT** (`<v Speaker N>`) in one job, plus a reusable `transcript_file_id` that cuts downstream chapters, highlight detection, filler removal, bleep censoring, translation, and burned-in captions to $0.01–$0.02/min.
- **Web tools:** [Transcribe audio](https://audo.ai/transcribe?utm_source=markdown&utm_medium=agent), [Interview transcription](https://audo.ai/interview-transcription?utm_source=markdown&utm_medium=agent), [Speaker diarization](https://audo.ai/speaker-diarization?utm_source=markdown&utm_medium=agent), [Podcast transcription](https://audo.ai/podcast-transcription?utm_source=markdown&utm_medium=agent), and [SRT file generator](https://audo.ai/srt-file-generator?utm_source=markdown&utm_medium=agent).

## Free API: one request, no key
`POST https://audo.ai/api/try/transcribe` transcribes the first minute of any audio or video file for free (3 times a day, no account or API key) and returns the transcript directly. Pick the output format with `format` (as a form field, JSON key, or `?format=` query parameter): `json` (the default), `txt`, `srt`, or `vtt`.

```bash
# SRT captions from a local audio file:
curl -F file=@interview.mp3 -F format=srt https://audo.ai/api/try/transcribe -D response.headers -o response.body

# Word-timed, speaker-labeled JSON from a local MP4 video:
curl -F file=@meeting.mp4 https://audo.ai/api/try/transcribe -D response.headers -o response.body

# Plain text in speaker turns, from a public HTTPS link:
curl -X POST "https://audo.ai/api/try/transcribe?format=txt" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com/interview.mp3"}' \
  -D response.headers -o response.body
```

Options are sent as form fields (or JSON keys) with the same names as the MCP tool's inputs:

- `language`: spoken language code such as `en`, `es`, or `ja` (default: detected automatically across 60 languages).
- `verbatim`: boolean (default `true`). Keeps filler words such as `um` and `uh` in the transcript so `remove_fillers` can cut them cleanly later.
- `speaker_labels`: boolean (default `true`). Labels who speaks when (`Speaker 1` through `Speaker 8`).

A request waits up to about 30 seconds and returns `200` with the transcript; longer jobs (or requests with `Prefer: respond-async`) return `202` with a `statusUrl` (`/api/try/jobs/<job_id>?t=<token>`). `Audo-Processed-Seconds` reports how much of the file was transcribed, and a free run cut short carries `Audo-Finish-Url` and `Audo-Finish-Price-Credits` (or `finishUrl` and `finishPriceCredits` in JSON) so a person can pay once from $5 to finish the whole recording.

To transcribe whole files up to 100 MB in one request with an API key from the [API page](https://audo.ai/keys?utm_source=markdown&utm_medium=agent), pass `Authorization: Bearer $AUDO_API_KEY`, `free=false`, and an approved positive `max_credits` cap (1 credit = 1 US cent):

```bash
curl -H "Authorization: Bearer $AUDO_API_KEY" \
  -H "Idempotency-Key: $(uuidgen)" \
  -F file=@interview.mp4 \
  -F format=json \
  -F speaker_labels=true \
  -F free=false \
  -F max_credits="$APPROVED_MAX_CREDITS" \
  https://audo.ai/api/try/transcribe -D response.headers -o response.body
```

For recordings above 100 MB (up to 10 GB), upload in parts via `POST /api/try/uploads` or use the [REST API](#rest-api) below. See [responses](https://audo.ai/docs/api?utm_source=markdown&utm_medium=agent#responses) and [whole files with an API key](https://audo.ai/docs/api?utm_source=markdown&utm_medium=agent#api-key).

## Output formats
### JSON

The canonical transcript document every other format is built from. Timestamps (`start_ms`, `end_ms`) are integer milliseconds from the start of the file. `speaker` and `speaker_count` are omitted when `speaker_labels` is `false` or speaker diarization could not run.

```json
{
  "version": 1,
  "language": "en",
  "duration_ms": 8200,
  "speaker_count": 2,
  "sentences": [
    {
      "start_ms": 0,
      "end_ms": 2100,
      "text": "Welcome back to the show.",
      "speaker": 1,
      "words": [
        { "text": "Welcome", "start_ms": 0, "end_ms": 420, "speaker": 1 },
        { "text": "back", "start_ms": 420, "end_ms": 700, "speaker": 1 },
        { "text": "to", "start_ms": 700, "end_ms": 820, "speaker": 1 },
        { "text": "the", "start_ms": 820, "end_ms": 960, "speaker": 1 },
        { "text": "show.", "start_ms": 960, "end_ms": 2100, "speaker": 1 }
      ]
    },
    {
      "start_ms": 2600,
      "end_ms": 4300,
      "text": "Thanks for having me.",
      "speaker": 2,
      "words": [
        { "text": "Thanks", "start_ms": 2600, "end_ms": 3000, "speaker": 2 },
        { "text": "for", "start_ms": 3000, "end_ms": 3150, "speaker": 2 },
        { "text": "having", "start_ms": 3150, "end_ms": 3600, "speaker": 2 },
        { "text": "me.", "start_ms": 3600, "end_ms": 4300, "speaker": 2 }
      ]
    }
  ]
}
```

### TXT

Readable speaker-turn paragraphs. When two or more speakers are detected, each turn starts with `Speaker N:`:

```text
Speaker 1: Welcome back to the show.

Speaker 2: Thanks for having me.
```

### SRT and VTT

Numbered SubRip (`SRT`) subtitles and `WebVTT` captions, where multi-speaker recordings wrap each cue in a `<v Speaker N>` voice span:

```text
1
00:00:00,000 --> 00:00:02,100
Welcome back to the show.

2
00:00:02,600 --> 00:00:04,300
Thanks for having me.
```

```text
WEBVTT

00:00:00.000 --> 00:00:02.100
<v Speaker 1>Welcome back to the show.

00:00:02.600 --> 00:00:04.300
<v Speaker 2>Thanks for having me.
```

## Speaker diarization & speaker labels
- **On by default:** `speaker_labels: true` runs automatically at no extra per-minute fee; pass `speaker_labels: false` when you do not want speaker turns.
- **Speaker numbering:** Speakers are numbered `1`, `2`, and so on in the order they first speak, up to **8** speakers. Labels identify voice turns rather than real-world names.
- **Clean turn boundaries:** A sentence never spans two speakers, so sentence-level cuts and captions align cleanly to speaker changes.
- **Graceful fallback:** The job's `result.speaker_labels` is `on` (with `speaker_count`), `off`, or `failed`. If diarization fails on degenerate audio, the word-timed transcript still succeeds in full and `result.warning` explains why.

## REST API: whole files up to 10 GB
For production backends, files over 100 MB (uploaded in parts over 16 MB up to 10 GB), and downstream `transcript_file_id` chaining, use `https://audo.ai/api/v1` with a bearer key from the [API page](https://audo.ai/keys?utm_source=markdown&utm_medium=agent) ([REST API reference](https://audo.ai/docs/api?utm_source=markdown&utm_medium=agent#rest-api)):

```bash
# 1. Create the file (returns upload.url up to 100 MB, or multipart partUrl/completeUrl over 16 MB).
curl -X POST https://audo.ai/api/v1/files \
  -H "Authorization: Bearer $AUDO_API_KEY" -H "Content-Type: application/json" \
  -d '{"filename": "interview.mp3", "sizeBytes": 48213000}'

# 2. After PUTting the bytes, wait until status is ready (prices shows exact credit quotes).
curl "https://audo.ai/api/v1/files/file_8Kc2QmP4?wait=30" -H "Authorization: Bearer $AUDO_API_KEY"

# 3. Start transcribe_audio (free: true runs the free part; omit free or pass free: false + max_credits for the whole file).
curl -X POST https://audo.ai/api/v1/jobs \
  -H "Authorization: Bearer $AUDO_API_KEY" -H "Content-Type: application/json" \
  -d '{"tool": "transcribe_audio", "file_id": "file_8Kc2QmP4", "idempotency_key": "<a new UUID>", "speaker_labels": true, "free": true}'

# 4. Long-poll the job until status is succeeded, failed, or cancelled.
curl "https://audo.ai/api/v1/jobs/<job_id>?wait=30" -H "Authorization: Bearer $AUDO_API_KEY"
```

A succeeded job lists all four generated files in `outputFileIds`—JSON first, followed by TXT, SRT, and VTT—and summarizes the transcript in `result`:

```json
{
  "language": "en",
  "duration_seconds": 1834.2,
  "word_count": 5120,
  "speaker_labels": "on",
  "speaker_count": 2,
  "transcript_file_id": "file_3Jd9RtY2",
  "excerpt": "Speaker 1: Welcome back to the show. …"
}
```

Fetch a 15-minute signed download URL for any output via `GET /api/v1/files/<file_id>/download`, or paginate through a time window of the transcript directly over HTTP without downloading the full JSON:

```bash
curl "https://audo.ai/api/v1/files/file_3Jd9RtY2/transcript?from=0&to=300&detail=full" \
  -H "Authorization: Bearer $AUDO_API_KEY"
```

This returns the sentences between `from` and `to` seconds with `startSeconds`, `endSeconds`, `text`, and `speaker` (plus per-word timings when `detail=full`), along with `nextFromSeconds` to fetch the next window.

## Batch-transcribe a folder & async polling
For quick local scripts under 100 MB, [download and inspect the reference Python client](https://audo.ai/examples/free-api.py?utm_source=markdown&utm_medium=agent) (`free-api.py`), which validates HTTP statuses, follows `202` job links on `audo.ai`, checks returned JSON/SRT/VTT/TXT payloads, and writes each recording's outputs into a fresh directory:

```bash
python3 free-api.py transcribe interview.mp3 ./transcripts/interview '{"format":"srt","speaker_labels":true}'

for file in ./recordings/*.{mp3,m4a,wav,mp4,mov}; do
  [ -e "$file" ] || continue
  name="$(basename "${file%.*}")"
  python3 free-api.py transcribe "$file" "./transcripts/$name" '{"format":"json","speaker_labels":true,"verbatim":true}'
done
```

For paid whole-episode batch jobs in production, send `Prefer: respond-async` so `/api/try/transcribe` returns `202` immediately with `statusUrl`, then long-poll with `&wait=25`:

```bash
for file in ./episodes/*.{mp3,m4a,mp4,mov}; do
  [ -e "$file" ] || continue
  name="$(basename "${file%.*}")"
  job="$(curl -fsS -X POST "https://audo.ai/api/try/transcribe?format=json" \
    -H "Authorization: Bearer $AUDO_API_KEY" \
    -H "Idempotency-Key: $(uuidgen)" \
    -H "Prefer: respond-async" \
    -F "file=@$file" -F speaker_labels=true -F free=false -F max_credits="$APPROVED_MAX_CREDITS")"
  status_url="$(printf '%s' "$job" | jq -r '.statusUrl')"
  while :; do
    poll="$(curl -fsS "${status_url}&wait=25")"
    state="$(printf '%s' "$poll" | jq -r '.status')"
    [ "$state" = "succeeded" ] && break
    [ "$state" = "failed" ] || [ "$state" = "cancelled" ] && { echo "Failed: $name" >&2; break; }
  done
  if [ "$state" = "succeeded" ]; then
    out_url="$(printf '%s' "$poll" | jq -r '.outputs[0].url')"
    curl -fsS "$out_url" -o "./transcripts/${name}.json"
  fi
done
```

Without an API key, one IP address can stream 2 uploads at once; an authenticated account can stream 4 uploads at once and run up to 3 jobs in parallel.

## MCP server: Claude Code, Cursor, Codex & ChatGPT
Add Audo's remote Streamable HTTP [MCP server](https://audo.ai/docs/mcp?utm_source=markdown&utm_medium=agent) (`https://audo.ai/mcp`) to Claude Code (`claude mcp add --transport http --scope user audo https://audo.ai/mcp`), Cursor, OpenAI Codex, Gemini CLI, VS Code, ChatGPT, or Claude ([setup guide](https://audo.ai/setup?utm_source=markdown&utm_medium=agent)). Your coding agent or assistant can then transcribe files and inspect long transcripts window-by-window without overflowing its context window:

1. `add_file` with a public `url` (or signed upload) returns a `file_id`.
2. `transcribe_audio` with `{"file_id": "file_8Kc2QmP4"}` runs transcription + speaker diarization and returns all four output formats alongside `result.transcript_file_id`.
3. `get_transcript` with `{"file_id": "file_3Jd9RtY2", "from_seconds": 0, "detail": "concise"}` reads speaker-labeled turns 5 minutes at a time by default (or pass `"detail": "full"` for word-level timestamps before calling `cut_audio`).

## Chain `transcript_file_id` into 7 downstream tools
Every `transcribe_audio` job on the REST API or MCP server returns `result.transcript_file_id`. Passing that ID to any of the seven transcript-aware tools below skips re-transcribing the recording and drops the per-minute rate to **$0.01–$0.02/min**:

| Tool | Name | Standalone | With `transcript_file_id` | Free | Output & billing notes |
|---|---|---|---|---|---|
| [Chapters and show notes](https://audo.ai/youtube-chapter-generator?utm_source=markdown&utm_medium=agent) | `write_show_notes` | $0.04/min | $0.01/min | First minute, 3 a day | YouTube chapters (`0:00` first, >= 10 s each), summary, Markdown show notes, key points, and title ideas; **$0.03 (`3` credits) minimum per paid run** |
| [Find highlights](https://audo.ai/ai-clip-finder?utm_source=markdown&utm_medium=agent) | `find_highlights` | $0.04/min | $0.01/min | First minute, 3 a day | Ranked, non-overlapping clip moments snapped to sentence boundaries (`title`, `reason`, `score`); **$0.03 (`3` credits) minimum per paid run** |
| Shorten transcript | `shorten_transcript` | $0.04/min | $0.01/min | First minute, 3 a day | Extractive `light`/`medium`/`strong` edit plan, waveform-safe `ranges` for `cut_audio`, `time_map`, and retimed JSON/SRT/VTT/TXT (web tool: [Meeting shortener](https://audo.ai/meeting-shortener?utm_source=markdown&utm_medium=agent)) |
| [Remove filler words](https://audo.ai/remove-filler-words?utm_source=markdown&utm_medium=agent) | `remove_fillers` | $0.03/min | $0.01/min | First minute, 3 a day | Cuts `um`/`uh` (8 languages) and exact repeats (`"the the"`) from audio and video in sync; `dry_run: true` lists selectable `cut_ids` |
| [Censor audio](https://audo.ai/bleep-censor?utm_source=markdown&utm_medium=agent) | `censor_audio` | $0.03/min | $0.01/min | First minute, 3 a day | Bleeps (`1 kHz`) or mutes swear words (7 languages), automatic AI PII (`pii: "all" | "contact"`, `+$0.01/min`, **$0.03** paid-run min), and custom `words` in place while keeping video untouched; `+$0.05/min` with `keep_background: true` |
| [Translate audio](https://audo.ai/audio-translator?utm_source=markdown&utm_medium=agent) | `translate_transcript` | $0.05/min | $0.02/min | First 2 minutes, 3 a day | Translates speech into any of the 60 supported languages while keeping speaker turns and cue timestamps (JSON, SRT, VTT, TXT) |
| [Add captions](https://audo.ai/auto-captions?utm_source=markdown&utm_medium=agent) | `add_captions` | $0.04/min | $0.01/min | First 30 seconds, 3 a day | Burns styled word-by-word captions onto video (36 presets; MP4, transparent overlay, SRT/VTT/ASS/TXT); **$0.02/min** with custom `captions.words`, or **$0.01/min** with `transcript_file_id` (`+$0.01/min` once with `auto_keywords`, `auto_emoji`, or `translate_to`) |

```bash
# 1. YouTube chapters + Markdown show notes ($0.01/min with transcript_file_id, $0.03 min per paid run):
curl -X POST https://audo.ai/api/v1/jobs \
  -H "Authorization: Bearer $AUDO_API_KEY" -H "Content-Type: application/json" \
  -d '{"tool": "write_show_notes", "file_id": "file_8Kc2QmP4", "transcript_file_id": "file_3Jd9RtY2", "idempotency_key": "<a new UUID>"}'

# 2. Up to 5 ranked highlight clips ($0.01/min with transcript_file_id, $0.03 min per paid run):
curl -X POST https://audo.ai/api/v1/jobs \
  -H "Authorization: Bearer $AUDO_API_KEY" -H "Content-Type: application/json" \
  -d '{"tool": "find_highlights", "file_id": "file_8Kc2QmP4", "transcript_file_id": "file_3Jd9RtY2", "count": 5, "idempotency_key": "<a new UUID>"}'

# 3. Cut filler words (um, uh, repeats) from audio or MP4 video ($0.01/min with transcript_file_id):
curl -X POST https://audo.ai/api/v1/jobs \
  -H "Authorization: Bearer $AUDO_API_KEY" -H "Content-Type: application/json" \
  -d '{"tool": "remove_fillers", "file_id": "file_8Kc2QmP4", "transcript_file_id": "file_3Jd9RtY2", "idempotency_key": "<a new UUID>"}'

# 4. Redact auto-detected PII (pii: "all", +$0.01/min, $0.03 min) and custom phrases ($0.01/min base with transcript_file_id):
curl -X POST https://audo.ai/api/v1/jobs \
  -H "Authorization: Bearer $AUDO_API_KEY" -H "Content-Type: application/json" \
  -d '{"tool": "censor_audio", "file_id": "file_8Kc2QmP4", "transcript_file_id": "file_3Jd9RtY2", "profanity": "strong", "pii": "all", "words": ["acme corp"], "idempotency_key": "<a new UUID>"}'

# 5. Burn styled word-by-word captions onto video ($0.01/min with transcript_file_id):
curl -X POST https://audo.ai/api/v1/jobs \
  -H "Authorization: Bearer $AUDO_API_KEY" -H "Content-Type: application/json" \
  -d '{"tool": "add_captions", "file_id": "file_8Kc2QmP4", "transcript_file_id": "file_3Jd9RtY2", "idempotency_key": "<a new UUID>"}'
```

The keyless `/api/try/<tool>` endpoints do not accept file IDs, so `transcript_file_id` chaining requires an API key on `/api/v1/jobs` (or `/api/v1/workflow-runs` for multi-step recipes such as `clips` and `meeting_shortener`) or the MCP server. On the website, the same workflows power [YouTube chapter generator](https://audo.ai/youtube-chapter-generator?utm_source=markdown&utm_medium=agent), [AI clip finder](https://audo.ai/ai-clip-finder?utm_source=markdown&utm_medium=agent), [Remove filler words](https://audo.ai/remove-filler-words?utm_source=markdown&utm_medium=agent), [Bleep censor](https://audo.ai/bleep-censor?utm_source=markdown&utm_medium=agent), and [Auto captions](https://audo.ai/auto-captions?utm_source=markdown&utm_medium=agent).

## Comparing Audo AI with Whisper API, Deepgram, AssemblyAI & pyannote
Developers evaluating speech-to-text APIs typically weigh upload caps on video files, whether speaker diarization and word timings require separate models, and how much media plumbing remains after the transcript comes back (rates and limits verified November 2026):

| Option | Max upload & video input | Speaker diarization & word timings | Pre-recorded price (Nov 2026) | Media editing, trials & MCP |
|---|---|---|---|---|
| **Audo AI (`transcribe_audio`)** | Up to **10 GB** & **10 hr** audio or raw `MP4`/`MOV`/`MKV`/`WebM`/`AVI` via multipart upload; zero pre-chunking or `ffmpeg` extraction | Up to **8 speakers** + millisecond word timings on by default across 60 languages; outputs JSON, TXT, SRT, and VTT | **$0.03/min ($1.80/hr)** PAYG from $5 (**$1.13–$1.44/hr** on $10–$100 packs); no multichannel multiplier; $0 on failure | Keyless 1-min Free API (`3x/day`), hosted MCP (`https://audo.ai/mcp`), and `transcript_file_id` chaining into filler cuts, bleep censoring, chapters, clips, and burned-in captions |
| **OpenAI Audio (`gpt-transcribe` / `whisper-1`)** | **25 MB** hard upload limit on `/v1/audio/transcriptions`; larger files or videos require client-side `ffmpeg` demuxing and chunk splitting | `gpt-transcribe` returns text + detected language only; word timings (`whisper-1`) and diarization (`gpt-4o-transcribe-diarize`) use legacy models scheduled for Feb 26, 2027 shutdown | `$0.0045/min` (`$0.27/hr`) on `gpt-transcribe`; `$0.003/min` (`gpt-4o-mini-transcribe`) to `$0.006/min` (`whisper-1` / `gpt-4o-transcribe-diarize`) | API key required; returns transcript text/JSON only, with no media cutting, redaction, burned-in captions, or audio/video MCP server |
| **Groq Whisper (`whisper-large-v3` / `turbo`)** | **25 MB** (free tier) / **100 MB** (dev tier); larger recordings require 16 kHz FLAC downsampling or overlapping chunk stitching | Word and segment timestamps (`verbose_json`), **no built-in speaker diarization**, and no native SRT/VTT subtitle output | `$0.04/hr` (`whisper-large-v3-turbo`) to `$0.111/hr` (`whisper-large-v3`), 10 s minimum billed per request | API key required; fast single-speaker text/JSON inference only, with no speaker turns or media editing |
| **Deepgram (`Nova-3` / `Nova-2`)** | **2 GB** request cap and 10-min processing timeout (`504`); docs recommend extracting audio streams from large video files first | Word timings + speaker diarization included on pre-recorded `Nova-3` (`+$0.0020/min` on streaming) across 60+ languages | `$0.0043/min` (`~$0.26/hr`) monolingual or `$0.0052/min` (`~$0.31/hr`) multilingual PAYG ($200 signup credit; multichannel bills per channel) | API key required; stateless JSON response (transcripts are not stored) with no filler cutting, bleep censoring, burned-in captions, or video stream-copy |
| **AssemblyAI (`Universal-3.5 Pro` / `Universal-2`)** | **2.2 GB** via `/v2/upload` (**5 GB** via URL), up to 10 hr | Utterance speaker labels + millisecond word timings across 18 (`Universal-3.5 Pro`) or 99 (`Universal-2`) languages | `$0.15–$0.21/hr` base + `$0.02/hr` standard diarization (`$0.17–$0.23/hr`; $50 signup credit; multichannel bills per channel) | API key required; transcript & LLM analysis stack (audio-only PII bleep add-on `+$0.05/hr`, but no filler cutting, media shortening, burned-in video captions, or MP4 stream-copy) |
| **Local `faster-whisper` + `pyannote.audio`** | Uncapped on local storage once you build multipart S3/R2 ingest, `ffmpeg` video demuxing, and 16 kHz mono WAV conversion | Requires orchestrating two separate models (`CTranslate2` Whisper + gated Hugging Face `pyannote/speaker-diarization-3.1` token) and custom word-to-turn alignment | Open-source weights, paid in hourly CUDA GPU instances or 15–60 s serverless GPU cold starts | Full DIY infrastructure for worker queues, subtitle formatting, and downstream FFmpeg audio/video editing |

- **When Deepgram or AssemblyAI wins:** You need sub-300 ms live WebSocket streaming for voice agents and telephony, or you process high-volume audio where raw transcript JSON at `$0.17–$0.31/hr` (billed per channel on stereo/multitrack files) is the final deliverable and you never edit or caption the underlying media.
- **When OpenAI or Groq Whisper API wins:** You already send short, single-speaker `<25 MB` audio clips inside an OpenAI or Groq app and only need plain text at `$0.04–$0.36/hr` without multi-speaker diarization, native SRT/VTT files, or large video uploads.
- **When self-hosted `faster-whisper` + `pyannote.audio` fits:** You keep CUDA GPUs busy 24/7 inside an air-gapped VPC and already maintain your own `ffmpeg` demuxers, Hugging Face model images, and word-speaker alignment code.
- **When Audo AI wins:** You accept user-uploaded audio or raw `MP4`/`MOV` videos from **25 MB up to 10 GB (10 hours)** without writing chunking or audio-extraction workers; want **8-speaker diarization + word timestamps + JSON/TXT/SRT/VTT** in one call (and can clean noisy room mics first with [Noise Removal & Audio Cleanup API](https://audo.ai/docs/noise-removal-api?utm_source=markdown&utm_medium=agent)); want a **zero-key `/api/try/transcribe` endpoint** and **hosted MCP server** for Claude Code, Cursor, and Codex; or want to pass `transcript_file_id` straight into **filler removal, profanity/PII censoring, meeting shortening, chapters, highlight clips, and burned-in video captions** without building a media pipeline.

## Related docs and guides

- [Free API and REST API reference](https://audo.ai/docs/api?utm_source=markdown&utm_medium=agent)
- [Noise Removal & Audio Cleanup API](https://audo.ai/docs/noise-removal-api?utm_source=markdown&utm_medium=agent)
- [Audio & Video MCP Server for Claude Code & Cursor](https://audo.ai/docs/mcp?utm_source=markdown&utm_medium=agent)
- [Audio Cleanup & Transcription in n8n](https://audo.ai/docs/n8n?utm_source=markdown&utm_medium=agent)
- [Audio and video tools for AI agents (Claude Code, Codex, Cursor & more)](https://audo.ai/guides/audio-and-video-tools-for-ai-agents?utm_source=markdown&utm_medium=agent)
- [How to transcribe an interview](https://audo.ai/guides/how-to-transcribe-an-interview?utm_source=markdown&utm_medium=agent)
- [How to get a transcript from a Zoom recording](https://audo.ai/guides/zoom-recording-transcript?utm_source=markdown&utm_medium=agent)
- [MacWhisper alternative with speaker labels](https://audo.ai/guides/macwhisper-alternative?utm_source=markdown&utm_medium=agent)
