IndicTelephony-Bench v1.0

An open benchmark for speech recognition on Indian telephone speech: 25,393 human-curated utterances (30.2 hours) in nine languages, recorded over live phone lines and released exactly as the line delivered them, at 8 kHz. Most of it is code-mixed, much of it is short, and the business vocabulary a voice agent acts on is tagged in the references.

At a glance

Utterances	25,393
Audio	30.18 hours, 8 kHz mono 16-bit PCM WAV, embedded in the parquet shards
Files	one parquet shard per language, under <Language>/test-*.parquet
Calls	320
Languages	Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Tamil, Telugu
Code-mixed utterances	70.4%
Shorter than 2 s	30.0%
Median duration	3.10 s
Utterances with tagged keywords	66.0%
Split	test only: this is an evaluation set, not training data

Languages

Language	Code	Utterances	Hours	Calls	Code-mixed %	< 2 s %	Median s
Bengali	bn	2,211	3.25	40	90.9	6.9	4.20
English	en	3,672	3.01	164	–	63.8	1.42
Gujarati	gu	2,230	3.90	54	87.6	10.1	5.16
Hindi	hi	2,837	3.35	89	76.1	41.8	2.76
Kannada	kn	3,549	3.92	85	79.4	27.2	2.66
Malayalam	ml	2,319	3.29	20	92.1	8.9	4.08
Marathi	mr	2,111	3.43	34	97.8	3.7	5.22
Tamil	ta	3,173	3.00	56	86.1	36.4	2.44
Telugu	te	3,291	3.02	76	60.7	40.0	2.34

Each language was recorded on a small number of lines, so compare systems within a language rather than languages with each other.

Loading

from datasets import load_dataset

ds = load_dataset("ConvoZenAI/indictelephony-bench", split="test")               # every language
ta = load_dataset("ConvoZenAI/indictelephony-bench", "Tamil", split="test")      # one language

row = ds[0]
row["audio"]          # {"array": ..., "sampling_rate": 8000, "path": ...}
row["transcription"]

The audio is 8 kHz because that is what a telephone line carries. If a model needs 16 kHz, resample at load time, for example ds.cast_column("audio", Audio(sampling_rate=16000)), and keep in mind that upsampling adds no information above 4 kHz.

Data fields

Field	Description
audio	the utterance, 8 kHz mono 16-bit PCM
utterance_id	stable identifier, e.g. call_0001_chunk0014
call_id	the call the utterance comes from; use it to group utterances, for example for call-level bootstrap intervals
language	base language code (bn, en, gu, hi, kn, ml, mr, ta, te)
language_tag	every language present, e.g. en-hi for Hindi mixed with English
is_code_mixed	true when more than one language is present
duration_sec	duration in seconds
domain	business scenario of the turn (see below)
transcription	the curated reference transcript
transcription_normalized	the same transcript after the released scoring normalizer: NFKC, lower-cased, apostrophes dropped, other punctuation to space, numbers written as digit sequences, and a short map of alternative spellings of the same word. This is the string WER and CER are actually computed against.
keywords	comma-separated domain-salient spans of the transcript (amounts, product names, places); empty if none

Transcription convention

• English words in Latin script; words of the Indian language in their native script.
• No translation and no transliteration.
• Numbers written as spoken.
• Fillers kept.
• No punctuation.

Score against the reference only after applying the released normalizer to both the reference and the hypothesis (see Evaluation). Raw string comparison charges systems for casing, punctuation and number formatting that are not recognition errors. For WER and CER the transcription_normalized column already carries the reference in that form, so you only have to normalize your own output; the normalizer is released, so the two sides are treated identically: https://github.com/ConvoZenAI/indictelephony-bench. The keywords column is stored as curated, so normalize it yourself before computing the keyword miss rate.

Domains

Domain	Value	Hours	Share of hours %	Utterances
Real estate	real_estate	7.77	25.7	5,404
Automotive	automotive	6.18	20.5	5,567
Other	other	3.98	13.2	4,267
Banking & finance	banking_finance	3.04	10.1	2,760
Insurance	insurance	2.74	9.1	1,878
Telecom	telecom	2.58	8.6	1,887
E-commerce	e_commerce	1.89	6.2	1,864
Education	education	1.15	3.8	1,058
Healthcare	healthcare	0.85	2.8	708

Shares are of audio, not of rows: median turn length varies about twofold between domains, so a domain can hold more utterances and less speech than another.

How it was built

• Recording. Pairs of speakers hold a call over a live SIP telephone connection, working through turns scripted across business scenarios. The turns are spoken to another person rather than read aloud alone at a microphone, so the pauses and backchannels are real; the channel, codec, band limit and turn timing are real too. Only the wording is prepared.
• Segmentation. Calls were trimmed of ring and IVR lead and split into single-speaker turns.
• Curation. Curators transcribed each turn while listening to the audio, with machine transcripts from several systems shown as optional drafts, system identities hidden, and instructions that the audio takes precedence over any draft. Curators also assigned the language tag, the domain and the keywords.
• Filtering. Utterances flagged as containing personal information were removed together with their audio, along with empty and unintelligible turns.

Evaluation

Scoring code: https://github.com/ConvoZenAI/indictelephony-bench (Apache 2.0) -- the normalizer, all five metrics and the call-level bootstrap.

Every metric is computed after one normalizer, applied identically to the reference and to every system's output: NFKC and lowercase, apostrophes removed, other punctuation to space, numbers canonicalised so spoken and written forms compare equal, a short explicit list of alternative spellings of the same word, and filler spellings collapsed. It deliberately does not transliterate between scripts, which would merge genuinely different graphemes and forgive real errors. transcription_normalized is the reference after that step, so scores can be reproduced without reimplementing it.

The reference metrics are WER, CER, a mixed error rate for code-mixed text (English tokens as words, Indian-language tokens as aksharas), a keyword miss rate over the tagged keywords (one minus keyword recall, with no insertion term), and a semantic WER that forgives only verified same-word differences. Rates are corpus-level (total errors over total reference units). Score an utterance a system returned nothing for as fully deleted, and exclude an utterance from a system's denominator only when the system does not support one of its languages.

Limitations

• Scripted speech. Results describe telephone-band recognition of business vocabulary, not spontaneous conversation.
• Clean recordings. The benchmark tests band limitation and code-mixing, not robustness to background noise.
• Few lines per language. Speaker and language are confounded, so differences between languages are not meaningful.
• One reference per utterance. The normalizer and semantic WER absorb systematic variation, but spelling variants outside the normalizer's list still count as errors under WER and CER.
• Draft-assisted curation. Curators saw machine drafts, including from a system developed by the dataset authors. Drafts were blinded and several were shown.

Licence

Released under Creative Commons Attribution 4.0 International (CC BY 4.0). You may share and adapt the data, including commercially, provided you give credit, link to the licence, and state whether you changed anything. The citation below satisfies the attribution requirement.

The licence covers copyright. It does not grant rights over the speakers' personal data, and it comes with no warranty: the references are human work and may contain errors.

Ethical use. The recordings are of consenting speakers. We ask that you do not attempt to identify them, and that you do not use the audio to clone or synthesise their voices. This is a request, not a licence condition.

Citation

@misc{indictelephonybench2026,
  title        = {IndicTelephony-Bench: An Open ASR Benchmark for Indian Telephone Speech},
  author       = {Priyadarshi, Prasoon and Abdul Azeez, Zaheer and {ConvoZen AI}},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/ConvoZenAI/indictelephony-bench}},
  note         = {25,393 utterances, 30.18 hours, nine languages, 8 kHz telephony}
}