whisper-medium-uzbek

whisper-medium-uzbek

Uzbek speech recognition model, fine-tuned from openai/whisper-medium (769M) on 753 hours of human-transcribed Uzbek speech. Output is Latin-script Uzbek.

Results

Word Error Rate (%), lower is better; best result per row in bold. Greedy decoding on held-out test splits.

Usage

# pip install "transformers>=4.40" librosa soundfile
import librosa
from transformers import pipeline

asr = pipeline(
    "automatic-speech-recognition",
    model="Mannon/whisper-medium-uzbek",
    chunk_length_s=30,        # long-form audio is windowed automatically
)

# load & resample to 16 kHz — decodes wav/flac without ffmpeg
speech, sr = librosa.load("audio.wav", sr=16000)
result = asr(
    {"array": speech, "sampling_rate": sr},
    generate_kwargs={"language": "uz", "task": "transcribe"},
)
print(result["text"])

Use language="uz" and chunk_length_s=30 for long audio.

Training

Dataset composition — public train splits of each source (before filtering):

Hours are computed from clip durations of each public release (UzbekVoice: estimated from the mean clip length of 4.33 s). After filtering, the final training set is 753 h (≈ 11.5 h, 1.5% fewer than the raw sum).

Pseudo-labels, scraped audio and all evaluation test sets were excluded from training.

<a id="attribution"></a>

Attribution (USC, CC BY 4.0). "Our product uses ISSAI and TUIT Uzbek Speech Corpus, which is available under a Creative Commons Attribution 4.0 International License." — Musaev et al., USC: An Open-Source Uzbek Speech Corpus SPECOM 2021, LNCS 12997, Springer (doi:10.1007/978-3-030-87802-3_40).

• Full fine-tune · AdamW · lr 1.2e-5 · weight decay 0.1 · effective batch 64 · 10% warmup, linear decay · up to 10 epochs · SpecAugment · fp16
• Hardware: 2× NVIDIA V100S 32 GB (DDP)
• Checkpoint selected by validation WER on a held-out USC split

Evaluation

Test splits of FLEURS, Common Voice 22.0, USC and FeruzaSpeech. Greedy decoding with no_repeat_ngram_size=4; references and hypotheses normalized with uzbek_text_norm.

Limitations

• Weakest on encyclopedic speech with rare named entities.
• Noisy, far-field and dialectal audio is less covered; Cyrillic output is not a target.

License

Apache-2.0 for the model weights. Each training dataset keeps its own license.

Citation

If you use this model, please cite:

@misc{mannon_whisper_medium_uzbek_2026,
  title        = {whisper-medium-uzbek: Uzbek Speech Recognition},
  author       = {Mannon},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/Mannon/whisper-medium-uzbek}}
}

Data sources

@inproceedings{musaev2021usc,
  title     = {USC: An Open-Source Uzbek Speech Corpus and Initial Speech Recognition Experiments},
  author    = {Musaev, Muhammadjon and Mussakhojayeva, Saida and Khujayorov, Ilyos and Khassanov, Yerbolat and Ochilov, Mannon and Varol, Huseyin Atakan},
  booktitle = {Speech and Computer (SPECOM 2021)},
  series    = {Lecture Notes in Computer Science},
  volume    = {12997},
  publisher = {Springer, Cham},
  year      = {2021},
  doi       = {10.1007/978-3-030-87802-3_40}
}
@misc{commonvoice22,
  title  = {Common Voice Corpus 22.0 (Uzbek)},
  author = {{Mozilla Foundation}},
  url    = {https://commonvoice.mozilla.org/en}
}
@article{conneau2022fleurs,
  title   = {FLEURS: Few-shot Learning Evaluation of Universal Representations of Speech},
  author  = {Conneau, Alexis and others},
  journal = {arXiv:2205.12446},
  year    = {2022}
}

Datasets: Common Voice 22.0 · google/fleurs · issai/uzbek_speech_corpus · k2speech/FeruzaSpeech · UzbekVoice.