CooledNik/logatra-benchmark

Logatra Benchmark

An independent speech-to-text benchmark for Indian languages and Hinglish, built by Nemukai, the team behind Logatra. Seven commercial STT engines are scored on a public split and on a fresh held-out split of 2026 Creative Commons audio that no model could have trained on.

This repository holds the audio, manifests and every raw provider response, so the whole leaderboard can be re-scored offline without an API key.

Leaderboard and per-case explorer	nemukai.com/labs/logatra-benchmark
Same report as a Space	CooledNik/logatra-benchmark
Code, harness and methodology (MIT)	github.com/Nemukai/logatra-benchmark
Logatra, the transcription product	logatra.com

Results

Every engine on the same clips: 324 from the public split and all 132 held-out chunks. Lower is better.

Engine	Public WER	Held-out WER	Held-out median	US$ per audio hour
Sarvam saaras:v3	32.3%	22.8%	10.6%	~0.54
Sarvam saaras:v4	26.1%	25.8%	13.4%	~0.54
Gemini 3.5 Flash	26.5%	22.2% *	11.7% *	~0.43
Sarvam saarika:v2.5	37.3%	19.7% *	0.0% *	~0.54
OpenAI gpt-audio	53.5%	28.5%	19.8%	~1.39
OpenAI whisper-1	78.3%	47.6% *	39.0% *	0.36
Mistral Voxtral Small 24B	293.8%	72.0%	26.6%	~0.36

* Contributed to the silver held-out references, so its held-out score is an upper bound (see below).

What stands out:

• Saaras v4 is the strongest engine on the public split and the fastest of the seven (0.6 s median per clip), but on fresh held-out audio saaras v3 still edges it by about three points. Neither wrote the held-out references, so that gap is a clean measurement.
• Voxtral falls into repetition loops on Indic audio: 46 of 896 outputs are runaway repetitions, which is why its mean WER is above 100% while its median is far lower.
• Whisper collapses on conversational Hindi: usable on read speech, close to unusable on real interview audio.

What is here

Path	Contents
manifests/*.jsonl	one JSON object per sample: id, relative audio_path, reference_text, language, code_mixed, source, license, split
audio/fleurs/	1,200 FLEURS test clips, 12 Indic languages, 16 kHz mono WAV
audio/mucs2021/	3,136 MUCS 2021 Hindi-English code-switched clips, 16 kHz mono WAV
audio/heldout_yt/	132 chunks (about 48 minutes) cut from 11 CC-BY YouTube videos published January to June 2026
results/<system>/<sample_id>.json	the raw cached response of every provider call: text, latency and the provider's own payload
report/data.json, report/rows.json	aggregated and per-sample metrics (WER, CER, script-agnostic WER, latency, cost)
manifests/heldout_yt.adjudications.json	every adjudication decision for the 36 held-out rows that failed the automatic consensus gate
assets/	the figures on this page and the Logatra logo

The dataset viewer shows three configs: heldout (silver references), fleurs and mucs2021. Audio paths in the manifests are relative to the manifests/ folder.

from huggingface_hub import snapshot_download

path = snapshot_download("CooledNik/logatra-benchmark", repo_type="dataset")
# then score offline with the harness from github.com/Nemukai/logatra-benchmark

How the held-out split was made

Eleven Hindi and Hinglish YouTube videos published between January and June 2026, each released by its uploader under the Creative Commons Attribution license, were cut into 132 chunks of about 22 seconds. They are long-form interviews, podcasts and talks: conversational speech with code-mixing, the kind of audio Logatra customers upload. Only the chunks are shipped, never the full videos.

References are silver, not gold

There is no human gold transcription of the held-out split yet. Each reference is the medoid of three systems' hypotheses under a script-agnostic consensus distance, and the 36 rows that failed the automatic gate were adjudicated case by case. The systems that contributed hypotheses (saarika, Gemini, Whisper) score against references they partly wrote, so their held-out scores are upper bounds; self_ref_rate in report/data.json gives the exposure per system (saarika 58%, Gemini 38%, Whisper 5%). Read the methodology in the GitHub README before citing held-out numbers.

What is deliberately not here

The Gramvaani Hindi telephony subset (OpenSLR 118, 2,917 clips) is licensed for academic use only, with commercial use by permission and no redistribution grant. Its audio, transcripts and the provider outputs derived from it are excluded. The manifests keep the Gramvaani rows with an empty reference_text so sample ids stay stable; the GitHub repository documents how to download OpenSLR 118 yourself and merge the references back in.

Nothing in this dataset comes from Logatra customers or private recordings.

Licensing

Each subset keeps the license of its source, and every manifest row carries its own license field.

Subset	Source	License
fleurs	google/fleurs, test split (Conneau et al., 2022)	CC-BY-4.0
mucs2021	OpenSLR 104, MUCS 2021 subtask 2 (Diwan et al., 2021)	CC-BY-SA-4.0
heldout_yt	11 YouTube videos whose uploaders chose the Creative Commons Attribution license; per-video credit in ATTRIBUTION.md	CC-BY-3.0
results/, report/	Nemukai's own API calls and measurements, derived from the audio above	same license as the underlying subset; the MUCS-derived outputs are CC-BY-SA-4.0

If you hold the rights to any of the held-out videos and want your material removed, open an issue on the GitHub repository and it will be pulled promptly.

Disclosure

Logatra is Nemukai's transcription product, so we have a stake in these numbers. The benchmark exists because we needed to choose a provider and did not trust vendor-reported results. Logatra runs Sarvam saaras:v4 in production as of October 2026, even though saaras v3 leads on the held-out split here. Every raw provider response is shipped so the scoring can be checked.

Citation

@misc{logatra-benchmark-2026,
  title  = {Logatra Benchmark: speech-to-text on Indian languages and Hinglish},
  author = {Singh, Priyansh Kumar},
  year   = {2026},
  publisher = {Nemukai},
  url    = {https://huggingface.co/datasets/CooledNik/logatra-benchmark}
}

Please also cite the source datasets: FLEURS (Conneau et al., 2022) and MUCS 2021 (Diwan et al., 2021).