RTTM — Rich Transcription Time Marked Format

RTTM file structure

Plain-text, line-based. Each line is one speaker turn, with 10 space-separated fields. The important ones: type (always "SPEAKER"), file_id, channel, start_time (seconds), duration (seconds), and speaker_id. Example:

SPEAKER meeting_001 1 12.45 3.20 <NA> <NA> speaker_A <NA> <NA>
SPEAKER meeting_001 1 15.80 5.10 <NA> <NA> speaker_B <NA> <NA>

RTTM vs SRT/VTT

SRT/VTT store transcribed text for subtitle display. RTTM stores speaker timing without text — just "speaker A talked from 12.45s to 15.65s." They're complementary: RTTM tells you who was talking when; SRT/VTT tells you what was said. Combine them and you get a speaker-attributed transcript.

When you'd use RTTM

Mostly in research and tool integration. If you're benchmarking a diarization system, RTTM is the standard input/output format and metric tools (dscore, md-eval) consume it directly. For consumer transcription, you'd typically convert RTTM to a more user-friendly format (DOCX with speaker labels, SRT with cue-level speaker annotations) before showing to end users — which is what we do.

In practice

We don't expose RTTM to end users — diarization output gets folded into the transcript directly as "Speaker 1:", "Speaker 2:" annotations. But internally, our diarization pipeline produces RTTM as an intermediate format before merging with Whisper's text segments by overlapping timestamps.

Related terms

• Diarization
• DER
• Speaker embedding

Last updated: September 21, 2026