Speaker diarization

Diarization answers "who spoke when." Speech Revolutions labels who spoke each segment and attaches a speaker to every word, so you can render a transcript as a back-and-forth conversation instead of one wall of text.

Turn it on

speaker_labels is on by default. diarize is a Deepgram-compatible alias for the same flag, so code ported from Deepgram works unchanged.

# speaker_labels defaults to True; shown explicitly here.
result = client.transcribe("meeting.mp3", speaker_labels=True)

# diarize is a Deepgram-compatible alias for speaker_labels
result = client.transcribe("meeting.mp3", diarize=True)

Speaker turns: .utterances

The most convenient view is result.utterances — an AssemblyAI-style list of contiguous speaker turns. Each utterance carries its speaker, its text, the start/end of the turn, and the words that make it up.

result = client.transcribe("meeting.mp3", speaker_labels=True)

for u in result.utterances:
    print(f"[{u.start:.1f}s] Speaker {u.speaker}: {u.text}")
# [0.5s] Speaker SPEAKER_0: Hi, thanks for joining.
# [3.2s] Speaker SPEAKER_1: Happy to be here.

Per-word speakers: .words

Every entry in result.words also carries a speaker alongside its timestamps. Utterances are simply runs of consecutive words with the same speaker, grouped for you — but the per-word labels are there when you need finer control (e.g. highlighting the current speaker word by word).

for w in result.words[:4]:
    print(w.speaker, w.word, w.start, w.end)
# SPEAKER_0 Hi 0.50 0.68
# SPEAKER_0 thanks 0.68 0.99

Render speaker turns

To display a conversation, iterate utterances and print a new block whenever the speaker changes. Since utterances are already grouped by turn, this is a direct loop — optionally mapping raw labels like SPEAKER_0 to friendly names.

NAMES = {"SPEAKER_0": "Host", "SPEAKER_1": "Guest"}

def format_ts(seconds: float) -> str:
    m, s = divmod(int(seconds), 60)
    return f"{m:02d}:{s:02d}"

for u in result.utterances:
    name = NAMES.get(u.speaker, u.speaker)
    print(f"{format_ts(u.start)}  {name}\n  {u.text}\n")

Migrating from Deepgram

Beyond the diarize alias, result.to_deepgram() returns a Deepgram pre-recorded response shape — including results.utterances with integer speaker indices — so existing parsing code keeps working while you migrate. See the migration playbook.