Transcripts that know who said what

Here's the whole feature in twenty-four seconds: the block of text, the automatic split, the naming, and the subtitle export.

Why does a two-person interview transcribe as one block?

Because speech-to-text answers one question: what was said. It has no obligation to answer the second one, who said it. For a single narrator that's the whole job. For anything with more than one voice in it, an interview, a panel, a podcast, an earnings call with a Q&A section, it leaves half the meaning on the floor.

Consider a line like "September is live, but it isn't decided." Said by the guest, that's a forecast. Said by the interviewer, it's a challenge. Same words, opposite meaning. When the transcript is a flat list of timestamped lines, you recover the difference by inference: reading around the line, or scrubbing the audio back and listening for the voice.

That's fine once. It doesn't scale to a ninety-minute recording you're trying to search.

What diarization actually does

Diarization clusters a recording by voice and tags each word with a speaker index. It's a separate job from transcription, and holding the two apart in your head makes the rest of this much clearer:

• Transcription answers what was said. It turns audio into words.
• Diarization answers who spoke when. It clusters the audio by voice and tags each word with a speaker.

Run both and every line arrives attributed. The property that makes this practical is that diarization is unsupervised: it works from the sound of the voices alone. Nothing to set up beforehand. No voice samples to enroll. No need to tell it how many people are in the room. That's why it can be on by default for every upload instead of being a setting you have to find.

The trade-off is that the clusters come back anonymous. The engine can tell two voices apart; it has no idea whose they are. So speakers arrive as Speaker 1 and Speaker 2, and putting real names on them is your job, once per recording.

Why utterances lie about who spoke

Speech engines hand back two views of the same audio, and picking the wrong one silently corrupts your speaker labels. Utterances are pre-grouped chunks, split on silence at a pause of roughly 0.8 seconds. Words are individually timed and individually tagged. Building the transcript from utterances is the obvious move, and it's the wrong one.

Here's the failure. A fast exchange has no 0.8-second pause in it. Two people talking over the top of each other land inside a single utterance, that utterance carries a single speaker label, and one person's words get quietly credited to the other. Nothing errors. The transcript just says the wrong thing, confidently, in exactly the passages where the conversation got interesting.

So we build lines from the word level instead. A new line starts wherever the speaker tag changes, which puts turn boundaries at the actual change of voice rather than at the pauses. Long uninterrupted turns are then split at sentence ends once they pass 320 characters, so a five-minute monologue stays a set of scannable, clickable rows instead of one wall of text.

Naming, and then correcting

Automatic separation gets you most of the way, and treating it as finished is where transcription tools tend to go wrong. It's a probabilistic pass over audio, so it will be imperfect on hard recordings. The design assumption here runs the other way: the automatic pass is a first draft, and everything in it is editable.

Name each speaker once

Type a real name over Speaker 1. It's saved with the transcript and applied everywhere: every line in the reader, the subtitle exports, and the view anyone you share the link with sees. Each speaker also carries their own accent colour, so a long transcript stays readable at a glance while you scroll past it.

Fix what the pass got wrong

The failure modes are predictable, and each one has a direct fix:

• A line attributed to the wrong person. Reassign it to a different speaker.
• A line where the speaker changes mid-sentence. Split it in two at the cursor, then give each half its own speaker.
• A speaker the pass never found. Add one by hand and assign lines to them.
• One person split across two labels. Give both labels the same name. Merging is just naming.
• A misheard word or a drifted timestamp. Edit the text or the time directly, or insert and delete lines.

A recording that separated cleanly needs two edits: two names. A messy one needs a few more. Neither case makes you start over.

Do the names follow into subtitles?

They do, and whether they're written into each caption is a toggle rather than a fixed decision about the format. Downloads come as SRT or WebVTT. The reason it's a choice is that the right answer genuinely differs by destination: a video editor cutting an interview wants to see who's talking, while a subtitle track burned onto the finished video usually doesn't want a name prefix on every line.

1
00:00:04,000 --> 00:00:08,600
Ana Reyes: Thanks for coming on. Let's start with the September meeting.

2
00:00:09,000 --> 00:00:14,400
Mark Feld: Happy to be here. September is live, but it is not decided.

Cues only store a start time, so end times have to be derived. A caption runs until the next one begins, sized to its own text length at about 15 characters per second of reading time, and held between 1.2 and 8 seconds. That upper cap is what stops a long silence from leaving a subtitle parked on screen for minutes.

The honest limitations

Three things are worth knowing before you rely on this for something that matters.

Separation is best-effort. Distinct voices on a clean recording separate well. Overlapping speech, similar-sounding voices, and several people sharing one microphone are the hard cases, and on those you should expect to correct some lines by hand. In our experience a panel recorded on one room mic is the worst case by a distance.

Speaker numbering is per file. Indexes are only stable within a single recording. The same guest across two uploads isn't recognised as the same person, so you name them in each. If you interview the same people weekly, that's a real cost, and it's worth knowing up front rather than discovering on episode nine.

Uploads only, for now. Speaker identification runs on uploaded audio and video files. Live stream transcripts and YouTube backfills are transcribed without speaker separation.

Try it on a recording you already have

Take an interview or a panel that's sitting on your disk and upload it. The transcript comes back separated by voice, you put two names on it, and you have something you can read, search, archive, and hand to an editor. Instead of a block of text you have to listen back to in order to understand.

The full list of what shipped is in the changelog, and the API reference covers reading cues and speaker names programmatically.