Model & data planning
How much training data do you need for speaker diarisation?
Updated 2026-08-01 · 4 min read
Short answer
For speaker diarisation, volume matters less than composition. Determining who spoke when in multi-party Indian-language audio, including overlapped speech. The corpus profile that works is per-speaker isolated channels with a mixed reference, genuine overlap preserved, turn-level ground truth. Start with a pilot sized to move diarisation error rate, overlap detection recall measurably, confirm the gain on held-out data recorded under deployment conditions, then scale the configuration that worked rather than scaling everything.
Key takeaways
- Success is measured on diarisation error rate, overlap detection recall, speaker-count accuracy.
- The most common failure is overlap edited out during recording
- Data profile: per-speaker isolated channels with a mixed reference, genuine overlap preserved, turn-level ground truth.
What the model actually needs
Determining who spoke when in multi-party Indian-language audio, including overlapped speech.
- Per-speaker isolated channels with a mixed reference
- Genuine overlap preserved
- Turn-level ground truth
Metrics that tell you when you have enough
Collect against a metric, not against a number of hours. When a pilot batch moves the metric and a second batch of the same profile moves it less, you are at the point where composition, not volume, is the constraint.
- Diarisation error rate
- Overlap detection recall
- Speaker-count accuracy
Common mistakes
- Overlap edited out during recording
- Single-channel-only capture leaving no reliable ground truth
A sensible collection sequence
| Phase | Volume | Purpose |
|---|---|---|
| Pilot | 10–20 hours | Validate format, acoustics and annotation against your pipeline |
| First production batch | 100–300 hours | Move the primary metric and expose composition gaps |
| Targeted top-up | 50–150 hours | Fill the specific dialects, conditions or edge cases the eval exposed |
| Evaluation set | 5–20 hours | Held-out, deployment-condition data never used for training |
Services that supply this data
This use case is normally served by conversational speech data, audio annotation. Most programmes combine two of them, because raw collection without matched annotation rarely moves an applied metric on its own.
Hold back an honest evaluation set
Reserve deployment-condition data that never enters training. Teams that evaluate on data recorded in the same sessions as their training data consistently overestimate real-world performance, then discover the gap after launch.
Frequently asked questions
What data profile suits speaker diarisation?
Per-speaker isolated channels with a mixed reference, Genuine overlap preserved, Turn-level ground truth
Which metrics should we track?
Diarisation error rate, Overlap detection recall, Speaker-count accuracy
What goes wrong most often?
Overlap edited out during recording
Can we start small?
Yes. A 10–20 hour pilot delivered in your ingest format is the standard first step, and it usually exposes format or annotation mismatches that would have been expensive at volume.
Related reading
Turn this into a dataset specification
Tell us the languages, speaker count and minutes. You get a written scope, a protocol and a fixed price within one working day.