Spaces:

---

hf-audio

/

open_asr_leaderboard

Like 1.53k

Running on CPU Upgrade

Hello, I would like to request support for Urdu (ur-PK) in the Open ASR Leaderboard multilingual evaluation and submit my open-source Urdu Streaming Zipformer 20M model for consideration.

Model: https://huggingface.co/mahwizzzz/Urdu-Streaming-Zipformer-20M
Source code: https://github.com/mwzkhalil/Urdu-Streaming-ASR

Model details:
Architecture: Streaming Zipformer Transducer
Parameters: Approximately 20 million
Inference: ONNX FP32 and INT8 using sherpa-onnx
Language: Urdu
Release: v1.2.0-rc1
License: CC BY 4.0 (model repository)

The released model reports a WER of 28.03% and CER of 11.96% on 299 FLEURS Urdu test utterances under its documented evaluation protocol. These are independently reported results, not yet leaderboard-verified.

I noticed that Urdu is not currently available in the multilingual leaderboard. I would like to propose adding FLEURS Urdu (ur_pk), potentially alongside other suitable held-out Urdu evaluation datasets.

I am willing to contribute a reproducible evaluation configuration with pinned model weights, ONNX inference dependencies, documented decoding settings, and WER/RTFx reporting.

Could you advise whether Urdu can be added to the multilingual benchmark and whether a contribution through the existing pull request process would be appropriate?

Thank you for supporting multilingual speech recognition research.