12 languages · 273.5 hours

Who can follow a conversation when everyone talks at once?

Three diarization systems scored on a language-balanced Indic corpus. Lower DER is better. DiariZen wins the headline. Nemotron misses more speech and confuses less. Community-1 loses the identity of who is speaking.

21,684 mono sessions, 24 kHz PCM
2-8 speakers natural overlap, backchannels, interrupts
12 languages 1,807 sessions each, from Assamese to Telugu
01

The ranking

Overall DER at collar 0. The bar under each score is the error mix: missed speech, speaker confusion, false alarm.

DiariZen is lowest overall because 38% of the corpus is two-speaker. From three speakers up, Nemotron is equal or better.

02

Listen to one tape

A 35.8 s English clip. Malini is the blue channel, Satvik is the orange channel. Overlap is both channels lit. Play, then click a turn or a lane.

0:00 / 0:35
03

Why the ranking looks like this

38%
Two-speaker sessions decide the headline

DiariZen is the two-speaker specialist (11.1% vs Nemotron 20.6% vs Community-1 16.2%). That slice is 38% of the corpus, which is why DiariZen wins overall.

Recall vs identity

Nemotron misses a lot of speech (21.3% miss, 0.41% FA). DiariZen and Community-1 recover more speech (miss about 10-11%) but then confuse speakers. Community-1's confusion term (21.9%) is the largest single error in the study.

Collars forgive edges, not names

Going from 0 ms to 500 ms drops overall DER by about 6-7 points for every system. Confusion stays flat. The ranking does not change with collar.

Many speakers flip the leader

On 5-8 speakers Nemotron is best (43.2%), then DiariZen (46.0%), then Community-1 (54.1%).

English is the one inversion

Nemotron's easiest language is English (24.4%). DiariZen is flatter across Indic languages and best on Kannada (25.4%). Only English inverts the ranking.

Overlap compresses the field

Under 5% overlap, DiariZen is clearly ahead (16.8%). In the 15-30% band the three systems compress: 34.0 / 35.2 / 43.9.

04

Full scoreboard

Rows rank by DER unless you sort another column. Lower is better.

DER vs collar

05

Every language has a winner

Lowest DER at collar 0. Click a language below to rerank the table. Click a column to sort by that metric instead.

DER by language at collar 0 ms

06

Harder rooms

Speaker count and overlap. Click a band to rerank by that slice.

07

What was scored

NIST RTTM speaker turns and a full-file UEM on every session. Components are percent of the same reference speech, so miss + FA + confusion = DER.

DER = (missed speech + false alarm + speaker confusion) / reference speech. Collar applied equally at the start and end of each reference turn. Overlapping speech is not skipped.

Hours by language

Sessions by number of speakers

Overlap occupies 26.9 h (9.8% of audio). Most sessions sit in the 5-15% overlap band. Event inventory: 405,153 lexical turns, 96,350 backchannels, 22,902 filled pauses.

08

Cite this page

Use either form.

Plain text 

Mandal, I. (2026, November). Who can follow a conversation when everyone talks at once? - Indic Edition. https://iamindrayudh.github.io/indic-diarization/

BibTeX 

@misc{mandal2026indic,
  author = {Mandal, Indrayudh},
  title  = {Who can follow a conversation when everyone talks at once? - Indic Edition},
  year   = {2026},
  month  = {November},
  url    = {https://iamindrayudh.github.io/indic-diarization/}
}