whistle-medium

Whistle

Whistle is a multilingual and crosslingual ASR model pretrained with weak phonetic supervision using IPA transcriptions generated by LanguageNet G2P models. Unlike self-supervised or grapheme-based approaches, Whistle leverages phoneme-level representations to enable better data efficiency, crosslingual generalization, and reduced catastrophic forgetting. Trained and evaluated on the CommonVoice-based CV-Lang10 benchmark, Whistle demonstrates superior performance on both seen and unseen languages under limited-data conditions.

Whistle was proposed in the paper Whistle: Data-Efficient Multilingual and Crosslingual Speech Recognition via Weakly Phonetic Supervision by Saierdaer Yusuyin et al from THU-SPMI. The original code repository can be found here.

Model details

Whistle is a Conformer based encoder model, and trained using the CTC (Connectionist Temporal Classification) approach. It was trained on ~4k hours of labelled speech data sourced from the publicly available CommonVoice_v11

Whistle checkpoints come in three configurations of varying model sizes. Including small (90 MB), medium (218 MB) and large (543 MB). And subword-based and wav2vec-based model of small size are also trained for comprison. The multilingual ASR model are trained on CV-lang10 data and then is evaluated on test dataset of corresponding language whitout fine-tuneing. All of the pre-trained checkpoints are available on the Hugging Face Hub. The checkpoints are summarised in the following table with links to the models on the Hub:

Evaluation

Results are reported in Phoneme Error Rate (PER%) and Word Error Rate (WER%).

Evaluation on Public CommonVoice_v11

• %PER | Model | Parameters | en | es | fr | it | ky | nl | ru | sv-SE | tr | tt | Avg. | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | | small | 90 MB | 8.02 | 3.37 | 5.68 | 4.04 | 8.29 | 5.77 | 6.05 | 18.07 | 8.32 | 8.53 | 7.61 | | medium | 218 MB | 6.70 | 2.63 | 4.53 | 3.12 | 5.95 | 3.95 | 4.61 | 14.81 | 6.04 | 8.47 | 6.08 | | large | 543 MB | 5.42 | 1.96 | 3.52 | 2.25 | 4.06 | 2.64 | 2.97 | 11.33 | 4.04 | 5.97 | 4.41 |

• %WER with 4-gram LM | Model | Model size | en | es | fr | it | ky | nl | ru | sv-SE | tr | tt | Avg. | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | ------ | | small | 90 MB | 10.76 | 8.68 | 16.01 | 9.98 | 1.02 | 7.32 | 1.59 | 6.14 | 7.63 | 7.30 | 7.64 | | medium | 218 MB | 9.83 | 7.82 | 14.94 | 9.04 | 0.91 | 6.57 | 1.65 | 5.65 | 7.27 | 7.37 | 7.10 | | Large | 543 MB | 8.80 | 7.02 | 14.02 | 8.16 | 0.94 | 6.22 | 1.46 | 5.06 | 7.05 | 6.92 | 6.56 |

More performance please ref to benchmark

Training Data

All of our multilingual ASR model are trained with 10 languages of cv-lang10, which has been processed in lang-process. But for English wav2vec-base model and multilingul wav2vec-base model, only audio are used to train the model. The language ID and training hours of the ten languages are in the following table.

BibTeX entry and citation info

@article{yusuyin2025whistle,
  title={Whistle: Data-efficient multilingual and crosslingual speech recognition via weakly phonetic supervision},
  author={Yusuyin, Saierdaer and Ma, Te and Huang, Hao and Zhao, Wenbo and Ou, Zhijian},
  journal={IEEE Transactions on Audio, Speech and Language Processing},
  year={2025},
  publisher={IEEE}
}

Community

If you encounter problems in use, you can directly raise Issues on the github page.