Datasets:
Submitting a system
The leaderboard exists to be argued with. If you build or sell a speech-to-text system that handles Hindi, Indian English, or code-mixed speech, submit it, including if you expect to beat us.
Start here: contact SquadStack. The scoring code is supplied on request to gate-approved evaluators; it is not in this repository. We aim to respond within three working days.
Principles
- You submit raw transcripts. We normalise and score. If each submitter normalised their own output, the leaderboard would measure normalisation effort as much as recognition quality. Do not pre-clean, do not post-correct, do not map scripts.
- One pipeline for every system. Reference-locked normalisation, per-turn alignment, deterministic error counting and one pinned LLM judge at temperature 0, with the same prompts for all systems.
- Identical decoding settings across every call. Tuning per slice is benchmark fitting, and if we detect it the submission is rejected with the reason published.
- Pin your model revision. Where a vendor ships silently behind a stable name, the run date is the real version and the row will say so.
- You touch your rows and nobody else's. A submission adds or replaces only its own system; every other row stays byte-identical.
- We re-run what we can re-run. Where we cannot (access, cost, a self-hosted model we have no hardware for) the row says "unverified" and names what we could not check.
1. Produce transcripts
Transcribe the lead-channel audio of the 863 calls with your system and return one raw transcript per
sample_id, exactly as your system emits it, including any punctuation and your own turn or segment
breaks. Declare the pinned model_version and the date of the run. If your API cannot accept 8 kHz,
upsample before sending and say so in the notes. Return an empty string where the system returned nothing:
such calls are scored, not skipped.
2. We score
We run the same pipeline used for every row: normalise your output to the reference conventions, align it to the reference turns, count substitutions, deletions and insertions, and have the pinned judge tag each error as meaning-changing or harmless. Do not substitute a different judge.
The board reports:
| Metric | Reported as |
|---|---|
| Semantic WER | rate: meaning-changing errors / reference words. The headline. |
| Raw WER | rate plus its substitution, deletion and insertion components |
| Error categories | counts of the judge's categories |
The three raw components (substitutions, deletions, insertions) sum to raw WER.
Final-transcript latency (FTR P80) is published for the locked 50-call set (see LEADERBOARD.md). Planned: per-call latency records and hallucination on silence.
3. Send it
Include:
- The raw transcripts (one per
sample_id) - Vendor, display name, pinned model version, run date, category, licence
- A link to your model's training-data disclosure
- A statement that this benchmark was not used in training at any stage, including as evaluation-driven model selection
- Confirmation that decoding settings were identical across all calls
- Anything you upsampled, chunked differently, or had to work around
4. Verification
We re-run your system where the endpoint is reachable and compare. A difference beyond the run-to-run noise of the judge (about 0.5 points of semantic WER) gets a note on your row. This is not an accusation; it is what makes the public number worth reading.
What gets a submission rejected
| Pre-normalised or post-corrected transcripts | The scoring path is not yours to control |
| Per-call or per-slice decoding settings | Benchmark fitting |
An unpinned model id (latest) |
The result cannot be reproduced |
| Self-judged semantic errors | The judge is pinned and ours, for the same reason the normaliser is |
| Any evidence the system was trained on this benchmark | Licence violation; the row is removed and the reason published |
Rejections are published with their reason. A leaderboard that rejects silently is not auditable.
Disputing a number
If you think one of our numbers is wrong, you do not need our code to say so. The results table carries
every raw transcript, the normalised text that was scored, and every error as a record with its verdict
and category. Recount them with your own code (every accuracy metric in this release is a count over
those records) and send us the discrepancy. We will treat a confirmed difference as our bug and log it as
an erratum. Disputing a verdict rather than a count needs the judge trace for that error, which we will
send you for the rows in question.
Adding a metric or a slice
Write to us before building anything. A new metric has to justify what it catches that the existing stack misses and define its denominator unambiguously. A new slice needs at least 40 calls or it is descriptive only.
Adding data
Not accepted. The corpus is fixed within a MAJOR version, which is what makes numbers comparable across submissions.