rahulsharmavishwakarma/vesper-indic-bench

Vesper Indic Bench

Per-clip transcripts and scores of Indic speech recognisers on identical held-out clips
Part of the Vesper project · 25 October 2026

Vesper Indic Bench is the evaluation behind Vesper-IndicEAR. Every system decodes exactly the same clips: the first 300 per language of each of 90 held-out evaluation sets across 12 languages (29,192 clips, 61.9 hours). The dataset publishes, for every clip, its source, its reference where the source licence allows, and each system's transcript and scores, so any number in the model card can be recomputed or re-scored with a different metric.

No audio is included. Clips are referenced by their source dataset, pinned revision and id, with the sha1 of the audio file.

Systems

System	Decoding	Language
Vesper-IndicEAR (automatic language)	CTC greedy, decode-time script rules	from its own language ID
Vesper-IndicEAR (language given)	same, with the clip's language as a hint	given
Vesper-IndicEAR (no script rules)	unconstrained CTC greedy	—
IndicConformer 600M multilingual (ai4bharat/indic-conformer-600m-multilingual@2a77d305)	CTC, repository code	given
SraVaani-1.0 (ARTPARK-IISc/SraVaani-1.0@f5dd5358)	TDT greedy, repository code	automatic; no Urdu

Each baseline runs its own repository's inference code at the pinned revision. SraVaani has no rows for Urdu and IndicConformer none for English, which they do not support.

Results

Mean script-normalised CER over the languages every system supports (Urdu excluded, as SraVaani has no Urdu); lower is better, best in bold. 300 clips per language per set; per-clip transcripts are in Vesper Indic Bench.

Eval family	Vesper-IndicEAR (auto)	Vesper-IndicEAR (lang. given)	IndicConformer 600M	SraVaani-1.0	languages
Kathbath clean, unseen speakers	21.0	20.9	3.4	3.2	10
Kathbath clean, test-known speakers	20.5	20.4	3.3	3.1	10
Kathbath noisy, unseen speakers	24.9	24.6	4.3	4.4	10
IndicVoices valid	24.3	23.9	7.7	7.4	10
Vaani test	24.8	24.1	10.9	8.4	10
IndicContext	32.4	31.6	5.2	5.3	7
MUCS 2021 (hi, bn)	43.5	42.9	49.0	49.3	2

Over 65 (set, language) pairs, Vesper-IndicEAR is lower than the better of the two baselines in 1 automatically and 1 with the language given.

Files

File	Rows	Contents
data/clips.parquet	29,192	one row per (set, clip)
data/transcripts.parquet	141,792	one row per (set, clip, system)
data/vesper_asr_benchmark_slice.parquet	1,584	Vesper-IndicEAR on the clips of Vesper's ASR benchmark it covers, keyed by that dataset's sample_id
summary.json		per system, set and language: CER, WER and their script-normalised forms

clips.parquet: set, key (source / language / split / id), lang, source, source_repo, source_revision, source_licence, split, seconds, audio_sha1, code_mixed, reference, reference_alternatives (the same sentence with English words in native script, where the source provides one), reference_withheld (why the reference is not published).

transcripts.parquet: set, key, lang, system, hypothesis, detected_language (Vesper-IndicEAR only), cer, wer, cer_script_normalised (per clip, in %).

Metrics

Text is normalised before scoring (lowercase, no punctuation, NFKC, native digits, and the folds of Malayalam chillus, Bengali khanda ta, the Marathi eyelash ra and Urdu heh, tatweel and short vowels). Script-normalised CER accepts an English word written either in Latin or in native script, scoring against the closest whole reference: IndicVoices and Vaani write English words in Latin, while IndicConformer and SraVaani write them in native script, and strict CER would count that as an error. It is the metric the model card compares systems with. The scorer is indicear.metrics in the code repository.

Sources and licences

Set family	Source	Reference text
Kathbath clean / noisy test (known and unseen speakers)	AI4Bharat IndicSUPERB, CC BY 4.0 (Hub tag; see below)	published
IndicVoices valid	ai4bharat/IndicVoices, CC BY 4.0	published
Vaani test	ARTPARK-IISc/Vaani-transcription-part, CC BY 4.0	published
IndicContext (with code-mixed subsets)	ai4bharat/IndicContextEval, CC BY 4.0	published
FLEURS dev	google/fleurs, CC BY 4.0	published
LibriSpeech test-clean / test-other	openslr/librispeech_asr, CC BY 4.0	published
VoxPopuli English test	facebook/voxpopuli, CC0	published
MUCS 2021 Hindi-English / Bengali-English test	OpenSLR SLR104, CC BY-SA 4.0	withheld (share-alike)
Svarah	ai4bharat/Svarah (gated; CC BY 4.0 only in its card text, no licence tag)	withheld, as in Vesper's ASR benchmark

Kathbath's licence is stated two ways: its Hub tag says CC BY 4.0 and its card text says CC0. Its read sentences come from the IndicCorp web-text corpus. The references are republished here under CC BY 4.0 with attribution to AI4Bharat.

To get the audio, load each source at the listed revision and match key; the code repository's indicear.sources and indicear.ingest do this for every source. This dataset (ids, transcripts, scores and published references) is CC BY 4.0; references keep their source's licence and attribution.

Speaker protocol

Vesper-IndicEAR never saw a speaker of these sets in training: every dev and evaluation speaker was removed from its training data. Other systems may have trained on some of these sources and speakers (Kathbath's test-known speakers are its training speakers by design), so compare across systems with that in mind.

Limitations

• 300 clips per language and set: enough to separate systems several points apart, not to rank close ones.
• Read and prompted speech dominate; conversational speech is limited to IndicVoices, Vaani and MUCS.
• Timings are not published; they came from one Apple silicon Mac and are not comparable across systems run on CPU and GPU.

Citation

@misc{vesper_indic_bench_2026,
  title  = {Vesper Indic Bench: per-clip transcripts of Indic speech recognisers on identical held-out clips},
  author = {Sharma, Rahul},
  year   = {2026},
  url    = {https://huggingface.co/datasets/rahulsharmavishwakarma/vesper-indic-bench}
}