Audio Transcriber — Deepgram Nova-3 Speech-to-Text

URL in → transcript out. Pass one or more direct audio/video file URLs. Get one clean JSON row per file with the full transcript. Zero setup — no API key required.

For AI agents

Call this actor via Apify MCP (call-actor) with actor id kaz_kakyo/audio-transcriber and minimal input:

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

Cost: $0.010 per audio minute zero-setup, $0.004 BYOK, no key needed for zero-setup.

Structured output (one dataset row per file):

• Always: transcript, durationSeconds, minutesBilled, language, confidence, model
• Opt-in: summary, speakerTranscript (diarization), srt, utterances, words
• Failures: { "type": "error", "url", "error" } rows — bad links never kill the batch

Typical incumbent Apify transcription actors land around $0.015–$0.030/min. DIY Deepgram signup needs an account and key; this actor runs zero-setup at $0.010/min, or $0.004/min with your own key.

Why this one

• Cheapest way to run Deepgram on Apify. Zero-setup is $0.01/min ($0.60/hour). Bring your own Deepgram key (free, $200 credit, no card) and the actor fee drops to $0.004/min — you pay Deepgram's wholesale ~$0.0043/min directly, about $0.50/hour.
• Nova-3, your choice of model. Pick the model (nova-3, nova-2, whisper-large), the language, and which extras to compute.
• Built for pipelines and agents. Feed it a list of URLs, get one clean JSON row per file. Chain it after any scraper that outputs media links, schedule it, or call it via MCP.
• HTTP-only and light. Deepgram fetches your URL directly — nothing downloads through the actor, so runs are fast with no proxy or compute surcharges.

What it does

Give it direct links to audio or video files (mp3, wav, m4a, flac, ogg, opus, mp4, mov, webm, mkv — up to 2 GB each). For every file you get:

Field	Always / opt-in	Description
transcript	always	Full text, smart-formatted, paragraph breaks
durationSeconds, minutesBilled	always	Audio length and what you were charged
language, confidence, model	always	Detected/requested language, overall confidence
summary	opt-in	Short abstract summary (English)
speakerTranscript	with diarization	Speaker 0: ... labeled dialogue
srt	opt-in	Ready-to-save subtitle file content
utterances	opt-in	Timestamped spoken segments with speakers
words	opt-in	Per-word timing and confidence

• Failed files become { "type": "error", "url", "error" } rows — one bad link never kills the batch, and you're never charged for failures.
• Skipped inputs (invalid URLs, duplicates, anything past the 500-file cap, or files left over after your run's charge limit) are recorded as rows too, so your input list always reconciles against the dataset.
• Silent or speech-free audio that decodes successfully is billed by duration like any other file — Deepgram processes the full recording either way.

Pricing

Event	Price	When
Audio minute (zero-setup)	$0.010	No key provided — transcription cost included
Audio minute (your Deepgram key)	$0.004	deepgramApiKey set — you pay Deepgram directly at cost
Actor start	$0.00005	Per run

Minutes are rounded up per file. A 90-second podcast clip = 2 minutes = $0.02 (zero-setup). A 1-hour meeting = $0.60 zero-setup, or ~$0.50 all-in with your own key.

Worth knowing: the bring-your-own-key discount pays for itself fast. Deepgram gives new accounts $200 of free credit with no card — that alone covers ~775 hours of Nova-3 pre-recorded transcription before you pay Deepgram anything.

Quick start

{
    "audioUrls": ["https://example.com/interview.mp3"],
    "diarize": true,
    "includeSrt": true
}

Podcast pipeline (schedule it): scrape episode URLs with any RSS/podcast actor, feed the enclosure URLs here, get transcripts + summaries into your dataset every morning.

Meeting archive: point it at your recording exports (any public or signed URL works), turn on diarization and utterances, index the output in your search tool.

Agent use: this actor is MCP-ready — an agent can pass a media URL and read back the transcript JSON in one call.

Input reference

• audioUrls (required) — direct file links. Not YouTube/TikTok page URLs; use a media-downloader actor first if you need those. Minimal call: only this field.
• deepgramApiKey — optional. Stored encrypted, sent only to api.deepgram.com, never logged.
• model — nova-3 (default), nova-2, whisper-large.
• language — BCP-47 code (en, es, zh, ...). Use multi with nova-3 for code-switched audio, or set detectLanguage and forget about it.
• diarize, smartFormat, paragraphs, summarize, includeSrt, includeUtterances, includeWords — feature toggles, all documented in the input form.
• keyterms — nova-3 keyterm prompting: pass product names, jargon, or speaker names to boost their recognition.

FAQ

Do I need a Deepgram account? No — zero-setup mode works immediately. Bring a key only if you want the lower rate.

What audio quality/languages are supported? Nova-3 supports 30+ languages and handles noisy, multi-speaker, real-world audio; whisper-large covers 90+ languages for rarer cases.

Is my audio stored? The actor never downloads or stores your media — Deepgram fetches the URL, returns JSON, and the actor keeps only the transcript in your dataset.

File limits? 2 GB per file, up to 500 files per run, ~10 minutes max processing per file (Deepgram's cap). Batches run 4 files at a time (2 for Whisper, which has stricter concurrency limits). Long runs are checkpointed — a platform migration resumes where it left off instead of re-billing finished files.

Why did my URL fail? It must be a direct, publicly reachable media file (or signed URL). Page URLs (YouTube, Spotify, Drive share pages) are not media files.

---

If this saved you time, a Store review on the actor page helps a solo dev. Hit a problem? Open an issue and I'll fix it fast.

Telemetry

Each run records one anonymous telemetry event to the developer's own storage: a salted hash of the caller account ID (never the raw ID), the run origin (e.g. web console vs API), and a timestamp. No inputs, results, or personal data are collected - this only measures adoption (distinct and returning users).