Feature · Diarization
Speaker diarization that tells every voice apart.
PyAI Hear marks each turn in a meeting, interview, podcast, app session or phone recording with the person who said it.
What PyAI gives you for speaker diarization
- PyAI speaker diarization returns text and anonymous speaker tags together in one Hear async job at $0.001 per audio minute, so you skip bolting pyannote onto plain transcription yourself.
- Each SRT and VTT caption cue names its speaker, so subtitles show who is talking and renaming Speaker 1 to a real person becomes a find and replace.
- A one-hour file finishes in well under a minute, running 92 to 247 times faster than real time with no platform fee on top.
- Finished jobs arrive by signed webhook, retried with backoff on timeouts and 5xx, or by polling the job id, and inputs stay up to 7 days while results stay up to 30 days.
What does speaker diarization output look like?
Choose a recording and press play to watch tags appear. Every line here is invented for the demo.
Pick a moment, then press play
How does PyAI find who spoke when in a file?
Telling speakers apart is one option you turn on for a Hear job. Labels, text, the charge and the webhook are saved together in one step, so a webhook never arrives without its result.
Submit the audio
A request to the PyAI transcription jobs API carries an audio link or upload plus the speaker option. Hear answers 202 with a job id marked queued and holds the input for up to 7 days.
Chunk or split
The worker decodes the file. One track is cut into overlapping chunks for parallel work, while a two-channel file is divided by channel.
Group the voices
For one track, the diarization stage clusters turns that sound alike. Speech recognition runs on each chunk at the same time.
Stitch and tag
Chunks are joined back in order and every segment receives a speaker tag, then SRT and VTT are rendered if you asked for them.
Deliver
A signed job-completed event reaches your webhook address, and fetching the job by its id serves text, segments, speakers and formats.
Each person was recorded on a separate channel
Pick channel mode. Tags map to channels, so similar sounding voices and crosstalk cannot be swapped.
Everyone shares one microphone or one track
Pick diarize mode. Hear clusters by voice, and someone on your side renames Speaker 1 or Speaker 2 to a real person. See how to transcribe an 8 kHz Hindi-English call with speaker labels.
Who needs speaker diarization software?
Speaker diarization software helps whenever someone reads, quotes or measures a talk later. People who run their own servers often ask which free tools to use. A hosted job skips that choice.
Captions with names
An episode with two hosts and a guest returns cue files where each caption shows its speaker, so swapping tags for names is a find and replace.
Quotable interviews
An hour of interview tape becomes turns with timestamps, so a reporter can pull every answer from the source without scrubbing audio.
Per person minutes
A product team adds speaker tags to recorded meetings so summaries can credit decisions and action items to whoever raised them.
Agent turns only
Reviewers filter an in-app or website agent session by speaker in the PyAI workspace, and Trace can score the agent side against your rules.
Clean training rows
Segments with start, end and speaker load into a warehouse as rows, ready for talk ratio or per speaker keyword counts.
Speaker diarization FAQ
Does the whisper-1 alias do speaker diarization?
No. The whisper-1 compatibility alias returns text without speaker tags. Open source users typically add pyannote as a separate diarization step. On PyAI, the whisper-1 alias maps to the sync endpoint and also returns no speakers. Diarization needs a Hear async job instead.
How is diarization different from speaker identification?
Diarization separates voices and gives each an anonymous tag, without knowing anyone's identity. Speaker identification compares a voice with enrolled samples to name a specific person. PyAI Hear performs diarization and channel separation only, so mapping Speaker 1 to a real name happens in your application.
Does GPT transcription support diarization on PyAI?
PyAI's OpenAI compatible sync endpoint accepts names like whisper-1 and returns JSON, plain text or detailed JSON without speakers. Speaker tags come only from the transcription jobs API with the speaker option set, which returns text, segments and a speakers list once the job completes and the webhook fires.
Can PyAI separate speakers during a live conversation?
No. PyAI speaker separation runs on finished recordings in Hear jobs, and the streaming socket returns partial and final text with no speaker field. Omni agent sessions can be saved as stereo recordings, and a Hear job in channel mode then tags each side once the conversation ends.
PyAI is voice infrastructure. It hands you the audio and transcript, and does not replace your own models or review tools.