India's 22 scheduled languages are spoken by over a billion people in conversations that mix languages, often within one turn, overlap freely, and take place over telephones, in meetings, and in media. Knowing who spoke when, the task of speaker diarization, is essential to speech applications for such conversation: voice agents conversing with several people at once, speaker attributed transcription of meetings and calls, contact center analytics, and media monitoring. We present a systematic study of speaker diarization across all 22 languages on IndicDiarBench and introduce Anuvani-Diarization, which adapts a self-supervised end-to-end segmentation model to Indian conversation with uncertainty aware silver supervision. Evaluated zero shot alongside nine recent diarization systems, it reaches a duration weighted diarization error rate (DER) of 16.78% on all 1,164 recordings, the lowest of all systems we ran and second only to the best published result (16.0%), and leads the three strongest rivals in 14 of 22 languages. It matches its parent on six international benchmarks. We also show that the benchmark's utterance level annotation convention is a hidden variable worth 1.9 to 4.1 DER points, depending on the system, and that a convention learned from other Indian transcriptions does not transfer.
Figures & tables
Figure 2: Anuvani-Diarization is the most accurate of all systems we ran on IndicDiarBench. Duration weighted DER on all 1,164 recordings (zero collar, overlap scored). † Joint transcription and diarization services reported by the benchmark’s authors ( Mehendale et al., 2026 ) ; all others were run under our protocol (Table 3 ).
L1 speakers
Bench
Overlap
DER (%), pooled
Language
Family
Script
(M), 2011
hours
Cond.
(%)
Ours
Best other
Δ
Assamese
IA
Assamese
15.3
1.5
N
13.9
22.43
23.53
DZ
− 1.10
Bengali
IA
Bengali
97.2
12.6
NFW
7.8
17.09
16.76
P2
+ 0.34
Bodo
TB
Devanagari
1.5
1.6
N
15.2
24.34
22.77
NM
+ 1.56
Dogri
IA
Devanagari
2.6
1.4
N
24.2
21.07
21.99
DZ
− 0.92
Gujarati
IA
Gujarati
55.5
11.1
NFW
7.6
13.08
13.65
P2
− 0.57
Table 1: The 22 scheduled languages and Anuvani-Diarization’s accuracy on each. Family (IA Indo-Aryan, Dr Dravidian, TB Tibeto-Burman, AA Austroasiatic), benchmark hours, recording conditions (N near field, F far field, W in-the-wild) and overlap are from Mehendale et al. (2026) ; mother tongue speakers from the 2011 Census ( Registrar General of India, 2018 ) . “Best other” is the most accurate of pyannoteAI Precision-2 (P2), DiariZen md-v2 (DZ) and NVIDIA Nemotron-3-Diarization (NM) on that language; the better of ours and the best other is in bold, and Δ is ours minus the best other, in percentage points, computed from the unrounded values. Bench hours are hours in the benchmark. Anuvani-Diarization leads in 14 of the 22 languages.
Figure 3: The benchmark’s references segment speech into utterances; diarizers segment it at pauses. Left: share of same speaker gaps shorter than 0.5 s on all 1,164 recordings. Right: pooled DER as same speaker gaps shorter than g are bridged ( y -axis from 12.5%); dashed: Anuvani-Diarization with a convention learned from Indian telephone transcripts, scored once (§ 5.1 ).
System
Raw
g∗
After
Δ
Ours (Anuvani-Diarization)
17.11
0.8
15.21
− 1.90
pyannoteAI Precision-2
17.58
0.8
15.55
− 2.03
DiariZen md-v2
18.31
1.0
15.40
− 2.91
NVIDIA Nemotron-3-Diarization
18.71
1.0
14.58
− 4.14
Word aligned public test sets (ours)
VoxConverse v0.3 test
9.32
1.0
10.00
+ 0.68
Table 2: Matching the reference convention moves every system, in opposite directions on different benchmarks. Pooled DER before (Raw) and after bridging same speaker gaps shorter than g∗ , the best threshold on the benchmark itself (a sensitivity analysis, not a system result); Δ is computed from the unrounded values. Bottom: on word aligned references, bridging at g=1.0 s raises DER.
Figure 4: Anuvani-Diarization. The Indic-adapted segmentation network combines a pruned WavLM-Large encoder with four Conformer blocks (blue) and predicts 16 powerset classes every 20 ms. The inset opens one Conformer block. Overlap-free speaker embeddings and VBx clustering link local speakers across windows; chips summarise the adaptation recipe.
Run under our protocol
Reported by Mehendale et al. (2026)
System
Pool.
DW
Miss
FA
Conf.
System
DW
Miss
FA
Conf.
Ours (Anuvani-Diarization)
17.11
16.78
8.91
3.55
4.65
Sarvam AI †
16.0
6.3
3.9
5.9
pyannoteAI Precision-2
17.58
17.18
9.70
3.06
4.82
AWS Transcribe †
23.5
13.1
3.1
7.4
DiariZen WavLM-Large s80 md-v2
18.31
17.99
11.07
3.03
4.22
Deepgram Nova-3 †
32.0
18.3
5.4
8.3
NVIDIA Nemotron-3-Diarization
18.71
18.38
13.07
2.63
3.01
Azure STT †
34.8
24.4
1.7
8.7
pyannote speaker-diarization-community-1
21.65
21.22
11.25
2.94
7.46
ElevenLabs Scribe †
35.0
13.6
6.2
15.3
Table 3: Full IndicDiarBench results (1,164 recordings; DER in %, zero collar, overlap scored). Left: run under our protocol, pooled (Pool.) and duration weighted (DW); ours in bold. Right: joint transcription and diarization services ( † ), as reported and named by the benchmark’s authors, who do not give model versions for all of them. Every system was evaluated zero shot; 4-spk: 4-speaker; Miss: missed speech; FA: false alarm; Conf.: confusion. b Inputs beyond the model’s maximum length are scored as missed.
Configuration
Pool.
DW
Miss
FA
Conf.
DiariZen WavLM-Large s80 md-v2 (parent)
18.31
17.99
11.07
3.03
4.22
+ Indic adaptation (Anuvani-Diarization)
17.11
16.78
8.91
3.55
4.65
+ convention learned from telephone transcripts z
22.74
22.43
2.89
15.29
4.56
Table 4: Indic adaptation helps; a convention learned from telephone transcripts does not transfer. DER (%) on all 1,164 recordings, each row evaluated zero shot. z Fitted on 302 SPRING-INX calls and selected on 35 others under a fixed protocol; scored once (Section 5.1 ).
Figure 5: Anuvani-Diarization leads in near field and far field meetings and at six of seven speaker counts. Pooled DER on all 1,164 recordings by the number of reference speakers (left; 2 to 8 speakers cover 1,157 recordings) and by recording condition (right; bars left to right: DiariZen md-v2, NVIDIA Nemotron-3-Diarization, pyannoteAI Precision-2, Anuvani-Diarization).
Subset
Rec.
Ours
P2
DZ
NM
Near field
781
19.05
19.93
20.25
20.83
Far field
173
18.64
19.48
20.03
19.48
In-the-wild
210
11.51
10.78
12.55
13.58
2 speakers
294
12.26
12.46
14.54
14.22
3 speakers
227
13.81
14.03
14.96
15.51
4 speakers
234
17.62
17.48
18.38
18.50
Table 5: Pooled DER (%) by recording condition (all 1,164 recordings) and by number of reference speakers (1,157 recordings with 2–8 speakers). Rec.: recordings; P2: pyannoteAI Precision-2; DZ: DiariZen md-v2; NM: NVIDIA Nemotron-3-Diarization. Best in bold.
DiariZen
Test set
Files
md-v2
Ours
VoxConverse v0.3 test
232
9.17
9.32
NOTSOFAR-1 eval a
129
18.26
17.41
MagicData-RAMC test
43
10.97
11.13
AMI test (SDM)
16
13.93
14.00
AISHELL-4 test
20
10.06
9.85
Table 6: The adapted network matches its parent beyond India. DER (%), zero collar, overlap scored, both with the parent model’s clustering ( Chung et al., 2020 ; Vinnikov et al., 2024 ; Yang et al., 2022 ; Carletta et al., 2005 ; Fu et al., 2021 ; Yu et al., 2022 ) . The better of the two is in bold. SDM: single distant microphone. a A single channel device per meeting.