Speaker Diarization & Transcription: Who Spoke When

Find out who spoke when in any audio or video file, and get a transcript with speaker names, subtitles and RTTM in one run. Paste links to your files (MP3, WAV, M4A, MP4…, or Google Drive and Dropbox file links) and get, for every file:

• 🕒 Who spoke when (speaker diarization): speaker turns with start and end times, plus the same turns in standard RTTM for diarization tools.
• 🗣️ Transcript with speaker labels: who said what and when, as JSON segments, readable text, SRT subtitles and WebVTT with <v Name> voice tags.
• 🏷️ Names and roles: speakers are named when the recording says who they are (or from your list of known names), with roles such as host, guest, interviewer, agent or customer.
• 📊 Talk-time stats per speaker, counted from the timestamps: talk seconds and share, turns, words per minute, questions asked, longest monologue, interruptions and long pauses.
• 🌍 Most languages: the transcript is written in the language spoken.

No API key, no account to connect, no setup. It works for podcasts, interviews, meetings, calls, lectures and videos.

What can you do with it?

• Podcast show notes: see who said what, pull quotes per guest and give every guest their talk time.
• Meeting minutes by person: a transcript that says who raised each point, ready to summarise or search.
• Interview research: code qualitative interviews by speaker without listening to every recording twice.
• Diarization models: RTTM for training data, or as a baseline to compare against your own model with pyannote or dscore.
• Subtitles with speaker names: SRT and WebVTT voice tags for video players, editors and accessibility.
• Talk-ratio analytics: talk share, monologues and interruptions per speaker for calls, meetings and coaching.

What files does it accept?

Input	Example	Supported
Direct audio or video file link	https://example.com/episodes/episode-42.mp3	✅ mp3, wav, m4a, aac, ogg, opus, webm, flac, mp4, mov…
Cloud storage or recording tool file URL	a public file URL from S3, another bucket or your recording tool	✅ if it opens without logging in
Google Drive file	https://drive.google.com/file/d/…/view	✅ shared as "Anyone with the link"
Dropbox file	https://www.dropbox.com/scl/fi/…?dl=0	✅
Folders, or pages that need a login		❌ share each file's own link

Telephone-quality (8 kHz) audio works fine. Files up to 4 hours are supported. By default each file is capped at 120 minutes; raise Max minutes per file (up to 240) for longer files, or lower it to cap the cost of long ones.

How much does it cost?

Pay per minute of audio. No monthly fee.

Event	Price (FREE/BRONZE)	Lower on paid plans
Audio minute processed (turns, RTTM, transcript, SRT/VTT, stats)	$0.009 / started minute	down to $0.008
Actor start	$0.00005 / run	—

A 60-minute podcast costs $0.54; a 10-minute interview costs $0.09. Minutes are counted per started minute processed. Nothing is charged for a file that fails.

Cap your spend: set a maximum cost per run. When the budget runs out, the file in progress is processed only up to what the budget covers and marked truncated.

Free plan: the AI runs on paid infrastructure, so free-plan runs include a 3-minute trial per run; upgrade to process full files.

How to get speaker diarization (who spoke when)

1. Click Try for free. The example input processes a 1.7-minute sample podcast with three speakers (recorded with AI voices).
2. Paste your own links into Audio or video links, one per line.
3. Optional: set the Number of speakers if you know it, and add Known speaker names (for example the host and guests).
4. Run, then open the Files, Speakers, Who spoke when or Transcripts view, or download JSON, CSV or Excel.

Input example

{
  "audioUrls": ["https://example.com/episodes/episode-42.mp3"],
  "expectedSpeakers": 3,
  "speakerNames": ["Emma", "Raj", "Lena"],
  "maxMinutesPerFile": 120
}

Output

One record per file. Output from the example input (shortened):

{
 "url": "https://api.apify.com/v2/key-value-stores/T5HtoeQYut1ufwkKC/records/sample-podcast.mp3",
 "source": "direct",
 "durationSeconds": 104.8,
 "processedSeconds": 104.8,
 "truncated": false,
 "language": "en-US",
 "speakerCount": 3,
 "speakers": [
  {"label": "S1", "name": "Emma", "role": "host", "displayName": "Emma", "talkSeconds": 41.0, "talkShare": 0.45, "turns": 6,
   "words": 120, "wordsPerMinute": 176, "questionsAsked": 4, "longestMonologueSeconds": 11.2},
  "…"
 ],
 "turns": [{"speaker": "S1", "start": 0.4, "end": 6.1}, "…"],
 "segments": [{"start": 0.4, "end": 6.1, "speaker": "S1", "text": "Welcome back to Build Notes…"}, "…"],
 "stats": {"talkSeconds": 90.1, "silenceSeconds": 0.0, "longPauses": 0, "interruptions": 0, "speakerChanges": 16, "speechShare": 0.86},
 "rttm": "SPEAKER sample-podcast-3f2a9c 1 0.400 5.700 <NA> <NA> S1 <NA> <NA>\n…",
 "srt": "1\n00:00:00,400 --> 00:00:06,100\nEmma: Welcome back…",
 "vtt": "WEBVTT\n\n00:00:00.400 --> 00:00:06.100\n<v Emma>Welcome back…",
 "text": "[00:00] Emma: Welcome back to Build Notes…",
 "processedAt": "2026-10-15T08:00:00+00:00"
}

Times (start, end) are seconds from the start of the file. A turn is one speaker's consecutive speech, merged across gaps of 1 second or less. Everything that must be exact (times, talk seconds, word and question counts, interruptions) is computed by code from the timestamps.

Output formats

Format	Field	Use it for
Speaker turns (JSON)	turns	who spoke when, in your own code
RTTM	rttm	diarization scoring and training tools (pyannote, dscore)
SRT	srt	subtitles with speaker names
WebVTT	vtt	web video subtitles with <v Name> voice tags
Plain text	text	reading, search and summaries
JSON segments	segments	transcript with times and speaker per segment

How accurate is it?

Measured on synthetic recordings with a known script (made with AI voices):

• 3-person podcast: 17 of 17 speaker turns attributed correctly in each of 3 runs; median start-time error 0.1 seconds.
• 3-person support call at telephone quality (8 kHz): 15 of 15 turns correct in each of 2 runs; median start-time error 0.14 to 0.2 seconds.
• Best for 2–6 distinct voices. On a 9-speaker court hearing where several similar voices take turns, voices were merged into fewer speakers. For panels and large meetings, treat speaker labels as approximate.
• Overlapping speech: each moment is given to one speaker.

How fast is it?

Measured on Apify with the default 512 MB memory: a 2-minute file in about 10 seconds, and a 30-minute recording in about 3 minutes (191 seconds, of which 2 minutes is preparing the audio). More memory gives the run more CPU: with 2 GB the same 30-minute file took about 1 minute (63 seconds). Long files are processed in about 10-minute parts cut at pauses, and each part is told which speakers were already found, so a voice keeps its label for the whole file. Files in one run are processed one after another.

Use it with AI agents (MCP)

https://mcp.apify.com?tools=bossless/speaker-diarization

Ask your assistant "who spoke when in this recording?", and it runs this actor and reads back the speaker turns and transcript.

Use the API from code

curl -X POST "https://api.apify.com/v2/acts/bossless~speaker-diarization/run-sync-get-dataset-items?token=YOUR_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"audioUrls": ["https://example.com/episodes/episode-42.mp3"], "expectedSpeakers": 2}'

from apify_client import ApifyClient

client = ApifyClient("YOUR_TOKEN")
run = client.actor("bossless/speaker-diarization").call(run_input={
    "audioUrls": ["https://example.com/episodes/episode-42.mp3"],
    "speakerNames": ["Emma", "Raj"],
})
for item in client.dataset(run["defaultDatasetId"]).iterate_items():
    for turn in item["turns"]:
        print(turn["speaker"], turn["start"], turn["end"])
    open("episode-42.rttm", "w").write(item["rttm"])

FAQ

Do I need an OpenAI, Deepgram, pyannote or other API key? No. Everything is included in the price.

Which languages work? Most major languages. The transcript is written in the language spoken, and language tells you which one was detected.

Can it identify speakers by name? Yes, when the name is said in the recording ("I'm Emma, and today…") or given in Known speaker names and the recording makes clear who is speaking. Otherwise speakers are labelled S1, S2…, with a role when it's clear (host, guest, interviewer…).

What about overlapping speech? Each moment is given to one speaker, so when two people talk at once, the shorter interjection may be attributed to the main speaker. Interruptions are still counted in stats.

What happens with a private or broken link? It's skipped with a clear message and not charged.

Is my data stored or used for training? Files are downloaded only for the run and deleted when it ends. Results are saved in your own Apify dataset. The AI model runs on Google Cloud (Vertex AI). Google's terms say it won't use this data to train or fine-tune its models; it may cache requests for up to 24 hours and log them for abuse monitoring.

How are minutes counted? Per started minute of audio actually processed, so a 61-second file is 2 minutes.

Is this an official product of any provider? No. This is an independent tool and is not affiliated with or endorsed by any podcast, meeting, cloud storage or AI provider.

Related tools

• Call Recording Transcription & AI Call Analysis: speaker-labelled call transcripts with an AI summary, sentiment, action items and a QA scorecard.
• Chinese Video Transcriber: Bilibili videos and any audio/video link to timestamped text, translation and SRT.