Audio Transcriber - Speech-to-Text API

You already have a public link to an audio file. Paste that HTTPS URL, run the Actor, and read one Dataset row with the transcript, speaker labels, timestamps, and SRT/VTT. You do not need your own speech-to-text API key. If you prefer to use your own Mistral account, you can bring your own key.

The Console prefill is a 24-second public-domain NASA recording of Neil Armstrong stepping onto the Moon (Wikimedia Commons Armstrong Small Step.ogg). It bills one started audio minute, and the transcript should contain "one small step". Keep it for a first look, or replace audioUrls with your own MP3, WAV, FLAC, OGG, or WEBM links. One bad URL does not cancel the rest of the batch.

After the run, open Key-Value Store RUN_SUMMARY for the overall status, billed minutes, and the next action. Then open the Dataset. The next useful run is another file or batch, from the API, MCP, a schedule, n8n, or Make.

Latest updates

19 October 2026: the first version accepts direct audio URLs and returns one status per file. Full update history.

30-second quickstart

1. Open the Actor and keep the default sample URL, or replace audioUrls with your public HTTPS audio links.
2. Leave diarization and subtitles on unless you only need plain text.
3. Run. Expect one Dataset row per URL with status, text, timed segments with speaker labels, and srt / vtt when enabled.

Example input:

{
  "audioUrls": [
    "https://upload.wikimedia.org/wikipedia/commons/d/dd/Armstrong_Small_Step.ogg"
  ],
  "diarize": true,
  "contextBias": ["Armstrong", "Apollo"],
  "includeSrt": true,
  "includeVtt": true,
  "maxDurationSeconds": 3600
}

What data you get

Field	Meaning
status	SUCCESS, NO_SPEECH, FAILED, or SKIPPED
text	Full transcript
speaker_transcript	Speaker-labelled plain text when diarization is on
segments	Timed sentence-level cues with a speaker label (speaker_1, speaker_2, ...) when diarization is on
srt / vtt	Deterministic subtitles from segments
duration_seconds / billed_minutes	Provider duration and started-minute billing
error_code	Explicit failure code when not successful

NO_SPEECH is not a transport failure. Invalid media, unreachable URLs, oversize or over-duration files, and budget stops never look like empty success.

Pricing

Primary event: audio-minute at $0.007 per started provider-observed audio minute (ceil(seconds / 60)). With your own Mistral key the event is audio-minute-byok at $0.004 instead, and Mistral bills your account for the transcription.

Workload	Approx. buyer charge (PPE only)
10-minute file	$0.07
60-minute file	$0.42

Also charged: synthetic apify-actor-start ($0.00005 per GB of run memory; one event at the default 1 GB). Platform usage (compute, transfer, storage) is included in these prices: nothing else is billed for the run. Failed items, duration-probe failures, over-duration rejects, and verified no-speech results are not charged. Free Apify plans cannot run this Actor, with or without your own key.

Set maxTotalChargeUsd on API or agent runs to bound the bill. The lowest accepted limit is $0.008: it covers the start events at any memory size plus one audio minute.

Use your own Mistral API key (optional)

You do not need a key: by default the Actor uses its included transcription. If your team already works with Mistral, or needs its own retention and training settings, you can run the Actor with your own key instead.

Set it up (about two minutes)

1. Sign in at console.mistral.ai. Free mode gives API access without a card. Add billing there if you need higher limits.
2. Open API Keys in the left sidebar and click Create new key. Give it a name and, ideally, an expiry date. Copy the key right away: Mistral shows it only once.
3. In this Actor's input, open Use your own Mistral key (optional) and paste the key into Your Mistral API key. For API, MCP, or scheduled runs, send it as mistralApiKey:

{
  "audioUrls": ["https://example.com/interview.mp3"],
  "mistralApiKey": "YOUR_MISTRAL_API_KEY"
}

4. Run. RUN_SUMMARY shows "transcriptionKey": "own" and "billingEvent": "audio-minute-byok", so you can confirm which key was used.

What changes

	Included (default)	Your own key
Setup	None	A Mistral API key
This Actor charges	audio-minute, $0.007 per started minute	audio-minute-byok, $0.004 per started minute
Mistral bills you	Nothing	The transcription itself, at your plan's rate (Mistral pricing)
Rate limits, retention, training settings	The publisher's Mistral account	Your Mistral account

If something goes wrong

• The key is checked with Mistral before any audio is downloaded. If Mistral does not recognize it (HTTP 401), the run stops as FAILED right away and RUN_SUMMARY.nextAction says so. If your account runs out of credit during the run (HTTP 402), the run stops the same way. No audio minutes are charged for the failed files, and files that already finished stay in the Dataset.
• The Actor never falls back to the included key when your key fails, so the price never changes silently.
• Rate limits on your account (HTTP 429) are retried with backoff. Free Mistral mode has low limits, so a large batch may be slow or end with UPSTREAM_RATE_LIMITED rows.
• Free Apify plans cannot run this Actor, with or without your own key.

Your key stays private. It is an Apify secret input: stored encrypted, sent only to api.mistral.ai, never logged, and never written to the Dataset, RUN_SUMMARY, or the billing ledger. Audio you send is processed under your Mistral account's terms, so check its privacy settings before sending sensitive recordings.

Diarization and context bias

Turn on diarize for interviews and meetings. Pass product names, people, or jargon in contextBias (up to 100 single words; the transcription API refuses entries with spaces or commas, so enter Neil and Armstrong as two entries). Context biasing is optimized for English. Other languages are experimental.

API / MCP / schedule

Direct MCP endpoint (authenticated execution):

https://mcp.apify.com?tools=automa-flow/audio-transcriber

Example agent ask: "Transcribe these public HTTPS audio URLs with diarization and return SRT."

When an agent should call it

• You already have direct public HTTPS audio file URLs
• You need a bounded transcript with status and spend limits
• You will set maxTotalChargeUsd and read RUN_SUMMARY, then Dataset rows

When not to call it

• You only have YouTube, TikTok, Spotify, or page URLs (not supported)
• You need real-time streaming, translation, summaries, or 90+ languages
• You cannot estimate audio length and spend before the call

Anonymous MCP discovery uses search-actors and fetch-actor-details only. Running this Actor always requires the caller's own Apify authentication and budget.

Schedule example: a nightly batch of new interview files with the same input shape. Use Apify webhooks on the run for downstream delivery. This Actor does not add its own webhook product.

Batch input

Send up to 100 URLs. Exact duplicate URLs are transcribed and charged once, then copied to every matching input_index; the copies show duplicate_of and billed_minutes 0. One bad URL does not cancel the rest of a valid batch.

Output schema

Dataset views: Transcripts, Subtitles (SRT and WebVTT per file), Errors, and All. Agents should read Key-Value Store RUN_SUMMARY first for overall status, counts, billed minutes, and nextAction, then fetch Dataset items.

Failure semantics

Code	Meaning
INVALID_URL	Non-HTTPS, SSRF-unsafe, or malformed URL
UNSUPPORTED_FORMAT	Not readable audio, no audio track, or a video file
DECODE_FAILED	The audio could not be decoded for transcription
URL_NOT_REACHABLE	Download failed
DURATION_PROBE_FAILED	Could not measure duration safely
MEDIA_TOO_LONG / MEDIA_TOO_LARGE	Over maxDurationSeconds or 500 MiB
BUDGET_INSUFFICIENT	Remaining charge limit too low
NO_SPEECH	Verified empty speech
UPSTREAM_*	Provider timeout, rate limit, or failure
DEADLINE_EXCEEDED	The run timeout arrived before this file finished; not charged
CANCELLED	Not started because the spending limit or a provider stop ended the run; not charged
BILLING_UNCERTAIN	The charge for this file could not be confirmed; its transcript is kept in BILLING_RECEIPT and the run is not replayed
PROCESSING_ERROR	Unexpected error for this file only; not charged, the rest of the batch continues
FREE_PLAN_REQUIRES_UPGRADE	Free Apify plans cannot run this Actor
PLAN_CHECK_FAILED (run status)	The platform's paid-plan signal could not be read; nothing was transcribed or charged

Each finished file is written to the Dataset right away, so a timeout or abort never loses a row you were charged for. Files the run timeout cut off are SKIPPED with DEADLINE_EXCEEDED, and RUN_SUMMARY.deadlineReached is true.

RUN_SUMMARY.chargeLimitReached is true when the spending limit stops further minutes. A restarted run reconciles the billing ledger with charged events and Dataset indexes. An unresolved charge is BILLING_UNCERTAIN and is not replayed.

Supported formats, languages and limits

• Formats: MP3, WAV, FLAC, OGG, WEBM (including browser recordings without a stored length), M4A, AAC. Audio-only files are accepted even when the server labels them video/webm or application/octet-stream. Video files are refused.
• Languages: 13 (en, zh, hi, es, ar, fr, pt, ru, de, ja, ko, it, nl) via the upstream transcription model
• Max duration: 60 minutes per file
• Max size: 500 MB
• Max URLs per run: 100

A language hint input is deferred. Current provider docs mark language as incompatible with the timestamp detail required for SRT and VTT.

Privacy

Audio is sent to Mistral for transcription through their official audio API. You are responsible for the rights and consents to process that audio. Do not use this Actor for covert recording or unlawful interception.

• The included publisher key is never exposed to users and is never logged. If you bring your own mistralApiKey, it is handled as described in Use your own Mistral API key.
• Raw audio is not kept in the Dataset or Key-Value Store. Temporary files are deleted after processing.
• Default Mistral retention is not zero retention. Zero Data Retention is specific to the publisher account and is not confirmed here. Training opt-out is a separate control.
• Transcript text is not written to normal INFO logs.

Limitations

• No YouTube, social, or webpage resolution. No browser and no proxies.
• No local Whisper, Batch API, translation, summaries, or chapters.
• No word-level timestamps: with speaker labels on, the transcription provider returns sentence-level segments only.
• Not a lowest-price claim, a highest-accuracy claim, or a HIPAA or GDPR certification.
• Audio URLs must be reachable from Apify's cloud without a login, cookie, or IP allow-list.

Technical notes

Implementation uses HTTP only against POST /v1/audio/transcriptions. Media duration is probed locally with ffprobe before spend. Compact MP3, OGG, WEBM, FLAC, and WAV files are uploaded as they are. M4A/AAC, files without a stored length, and heavy files (above about 256 kbps, such as CD-quality WAV) are first converted to 16 kHz mono FLAC, the rate the speech model works at, which also makes the upload several times smaller. Billing always uses the provider-measured duration. Billing charges exact started minutes from usage.prompt_audio_seconds after a verified non-empty transcript.