Speaker Diarization — who spoke when, with meeting notes

Send audio or video file URLs and get back the conversation split by speaker: the transcript as speaker turns, per-speaker talk-time statistics, speaker-labelled SRT/VTT subtitles, and ready-to-paste Markdown meeting notes. Everything runs inside the Actor — Whisper for the words, voice-embedding clustering for the speakers, all on CPU. No API key, no external service, no per-provider markup.

Model note (Sep 2026): the default model is now tiny (fastest, cheapest); choose base for accents, other languages or noisy audio. small was retired — a request for it runs base.

What you get per file

• meeting_notes — Markdown you can paste straight into a doc: a participants table (talk time, share, turns, words) plus the turn-by-turn transcript with timestamps.
• turns — consecutive same-speaker speech grouped into turns: {speaker, start, end, text}.
• segments — fine-grained timestamped segments with speaker labels.
• speakers — per-speaker stats, ordered by talk time.
• srt / vtt — subtitles with SPEAKER_N: prefixes (VTT uses proper <v SPEAKER_N> voice tags, so capable players colour each speaker).
• text, detected language, duration_seconds, billed_minutes.

Speakers

Leave Number of speakers at 0 and the Actor detects how many voices are present (up to Max speakers). If you know the count — a 1-on-1 interview is 2, your weekly stand-up is 5 — set it explicitly: known counts are always more accurate than detection.

Speakers are labelled SPEAKER_1, SPEAKER_2, … in order of first appearance.

Pricing — you only pay for delivered results

• $0.005 per file that delivers a speaker-labelled transcript, plus
• $0.007 per started minute of audio processed.

A file that fails to download, cannot be decoded, or contains no recognizable speech is never billed — not even the per-file fee. max_minutes_per_file is your hard cost ceiling per file: max spend = $0.005 + $0.007 × that number.

Sources and limits

• Direct media file URLs: mp3, m4a, aac, ogg, opus, wav, flac, mp4, webm, …
• Up to 25 URLs per run, 250 MB per file.
• Social-platform pages (YouTube, TikTok, …) are not supported — send direct media files you have the right to process.
• tiny and base models are baked into the image, so no run spends time downloading weights. (small was retired on 2026-09-04 — it ran 3× slower than base for a marginal gain; inputs that still say small run base.)
• Long recordings are processed up to max_minutes_per_file (default 30, max 480) and flagged truncated. If a run approaches its own timeout, the Actor delivers what it has — the partial transcript, or the speaker-labelled prefix if the clock runs out during diarization — and bills only the minutes actually delivered.

Tips

• Meetings & interviews: keep VAD on, set num_speakers if you know it.
• Raise the default run timeout if you raise max_minutes_per_file far above the default — a 4-hour recording needs more than the default hour.
• Accuracy is best with one voice per channel-mixed recording of reasonable quality; heavily overlapping speech is attributed to the dominant voice.
• Low-bandwidth phone recordings (8 kHz call audio) of similar-sounding voices may be under-separated in auto-detect mode — voice embeddings lose most of their contrast there. Set num_speakers explicitly for call recordings; meeting/podcast-quality audio is unaffected.