ONNX Model Zoo

karim1993/zipvoice-sr-finetuned

By karim1993 · ↓ 3 · ♥ 0 · Updated 2026-03-22

What's in this repo

This page lists the files in karim1993/zipvoice-sr-finetuned, including fm_decoder.onnx, fm_decoder_int8.onnx, text_encoder.onnx and more.

  • 4 ONNX models (.onnx files)
  • 15 files, including 4 .onnx, 3 .json, 2 .pt, 2 .safetensors, 2 .txt, 1 .md
  • Last updated 2026-03-22

ONNX files

File Size Download
zipvoice_distill_sr/fm_decoder.onnx — Download
zipvoice_distill_sr/fm_decoder_int8.onnx — Download
zipvoice_distill_sr/text_encoder.onnx — Download
zipvoice_distill_sr/text_encoder_int8.onnx — Download

How to use

These ONNX files can be loaded with ONNX Runtime. To inspect a model graph, upload the file to the Netron viewer. Download links point to the original files on Hugging Face.

README

ZipVoice Multilingual Finetuning

Hugging Face Space   Hugging Face Serbian Model   Hugging Face Arabic Model

Fine-tuning and inference setup for multilingual ZipVoice models, focused on:

  • Arabic (ar)
  • Serbian (sr)

Base model used in this project: ZipVoice.

Attribution

  • GitHub: https://github.com/k2-fsa/ZipVoice

Hugging Face Model Links (Quick Access)

Demo page:

Arabic model page:

Serbian model page:

Training Datasets

Serbian

Arabic

Evaluation Metrics

Serbian (sr)

Metric Count Mean Median Std Min Max
wer 92 0.17 0.08 0.24 0.00 1.00
cer 92 0.10 0.02 0.25 0.00 2.00
wav_seconds 92 4.18 3.63 2.92 0.24 22.98
wavlm_sim 92 0.63 0.69 0.15 0.01 0.83

Arabic (ar)

Metric Count Mean Median Std Min Max
wer 100 0.14 0.00 0.20 0.00 1.00
cer 100 0.05 0.00 0.14 0.00 1.30
wav_seconds 100 3.39 2.66 2.20 0.62 12.48
wavlm_sim 100 0.45 0.48 0.14 0.01 0.69

Arabic WER caveat: Whisper does not output Arabic diacritics (علامات التشكيل / الحركات), so diacritic mismatches are not reflected in WER. ◌َ فَتْحَة | ◌ُ ضَمَّة | ◌ِ كَسْرَة | ◌ْ سُكُون | ◌ّ شَدَّة | ◌ً تَنْوِينُ الفَتْح | ◌ٌ تَنْوِينُ الضَّم | ◌ٍ تَنْوِينُ الكَسْر | ◌ٓ مَدَّة | ◌ٰ أَلِف خَنْجَرِيَّة

Training Summary

Reported training time:

  • 2 days for each language
  • 1 extra day for each distilled model

Arabic Training Data

  • total_rows: 153666
  • total_duration: 395.46 hours

Serbian Training Data

  • total_rows: 92177
  • total_duration: 280.87 hours

Gradio Inference

Main app:

  • gradio_infer_fixed.py

Local run:

python3 -m pip install -r requirements.txt
python3 -m gradio_infer_fixed

System packages required:

  • ffmpeg
  • espeak-ng

Notes For Hugging Face Spaces

  • This README includes Spaces metadata in the front matter (sdk: gradio, app_file: gradio_infer_fixed.py).
  • The app is configured for Space deployment and can auto-download model artifacts/source when enabled by environment variables.
  • Arabic model page (again): karim1993/zipvoice-ar-finetuned
  • Serbian model page (again): karim1993/zipvoice-sr-finetuned

Model info

text-to-speech
onnx zipvoice text-to-speech sr license:apache-2.0 region:us
↓ 3 ♥ 0

Created: 2026-02-17

Updated: 2026-03-22

View on Hugging Face

Files (15)

.gitattributes
README.md
config.json
zipvoice_distill_sr/checkpoint-2000.pt
zipvoice_distill_sr/checkpoint-2000.safetensors
zipvoice_distill_sr/fm_decoder.onnx ONNX
zipvoice_distill_sr/fm_decoder_int8.onnx ONNX
zipvoice_distill_sr/model.json
zipvoice_distill_sr/text_encoder.onnx ONNX
zipvoice_distill_sr/text_encoder_int8.onnx ONNX
zipvoice_distill_sr/tokens.txt
zipvoice_sr/epoch-15-avg-4.pt
zipvoice_sr/epoch-15-avg-4.safetensors
zipvoice_sr/model.json
zipvoice_sr/tokens.txt

Related models

More from karim1993

More text-to-speech models

Browse all text-to-speech models

Popular models