GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Mannon
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Mannon
Model Tree
openai/whisper-medium
Input Modalities
Output Modalities
Supported Functionality
Word Error Rate (%), lower is better. This model: whisper-medium. Greedy = num_beams=1; beam 16 = num_beams=16 (best per model in bold).
| Test set | small · greedy | small · beam 16 | medium · greedy | medium · beam 16 |
|---|---|---|---|---|
| Parameters | 244M | 769M | ||
| FLEURS | 16.96 | 16.21 | 14.11 | 13.01 |
| Common Voice 22.0 | 9.58 | 9.14 | 7.69 | 7.01 |
| USC | 11.81 | 11.22 | 9.7 | 9.1 |
| FeruzaSpeech | 7.95 | 7.31 | 7.41 | 6.96 |
python
# pip install "transformers>=4.40" librosa soundfileimport librosafrom 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 ffmpegspeech, 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.
| Source | Version | License | Description |
|---|---|---|---|
| Common Voice | Corpus 22.0 (uz) | CC0-1.0 | crowd-sourced read speech |
| FLEURS | uz | CC-BY-4.0 | read speech, train split only |
| Uzbek Speech Corpus (USC) | v1 | CC-BY-4.0 | read speech (ISSAI & TUIT) |
| UzbekVoice | — | Apache-2.0 | crowd-sourced read speech |
Dataset composition — public train splits of each source:
| Source | Split | Hours | Utterances | Share |
|---|---|---|---|---|
| UzbekVoice (filtered) | train | ≈ 602.8 | 501,330 | 78.8% |
| USC | train | 96.4 | 100,767 | 12.6% |
| Common Voice 22.0 | train | 55.2 | 48,733 | 7.2% |
| FLEURS | train | 10.1 | 2,943 | 1.3% |
| Total | ≈ 764.5 | 653,773 | 100% |
Hours are computed from clip durations of each public release (UzbekVoice: estimated from the mean clip length of 4.33 s). Together the public train splits of these sources total ≈ 764.5 h.
Pseudo-labels, scraped audio and all evaluation test sets were excluded from training.
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).
Test splits of FLEURS, Common Voice 22.0, USC and FeruzaSpeech. Greedy decoding and beam search (num_beams=16), both with no_repeat_ngram_size=4; references and hypotheses normalized with uzbek_text_norm.
Apache-2.0 for the model weights. Each training dataset keeps its own license.
If you use this model, please cite:
bibtex
@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}}}bibtex
@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.
# pip install "transformers>=4.40" librosa soundfileimport librosafrom 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 ffmpegspeech, sr = librosa.load("audio.wav", sr=16000)result = asr( {"array": speech, "sampling_rate": sr}, generate_kwargs={"language": "uz", "task": "transcribe"},)print(result["text"])@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}}}@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}}