sherpa-onnx-vits-rasa-13

ONNX export of ai4bharat/vits_rasa_13 packaged for use with sherpa-onnx.

Model Overview

• 1024 speakers
• 32 emotions
• 13 Indic languages
• Single ONNX file
• Adds emotion_id as a 7th input to the standard sherpa-onnx VITS runner signature. (Stock sherpa-onnx does not know about this input. See requirements)

Requirements

• sherpa-onnx built with emotion-input support. This model exposes emotion_id as a required ONNX input. PR in sherpa-onnx (#3849). On earlier builds you will see unexpected input 'emotion_id' at load time.

Files

File	Purpose
model.onnx	fp32 VITS graph, ~150 MB
tokens.txt	Character frontend, one <char> <id> per line

Metadata

Key	Value
model_type	vits
frontend	characters
add_blank	1
n_speakers	1024
num_emotions	32
sample_rate	22050
language	multilingual-indic
punctuation	(empty)

Valid runtime ranges:

• speaker_id: [0, 1024)
• emotion_id: [0, 32)

Usage

sherpa-onnx-offline-tts \
  --vits-model=./model.onnx \
  --vits-tokens=./tokens.txt \
  --sid=0 \
  --emotion-id=0 \
  --output-filename=./out.wav \
  "வணக்கம், நீங்கள் எப்படி இருக்கிறீர்கள்"

Notes:

• sherpa-onnx splits input on ., ?, ! regardless of this model's punctuation ONNX metadata. If your input ends with sentence-ending punctuation, sherpa generates a short trailing sentence that may synthesize as noise. Workarounds: strip trailing punctuation, or use --vits-noise-scale-w to try to reduce it. Tracking issue: TBD.

Reproducing the export

Export tooling lives in a separate repository: github.com/matiaslin/sherpa-onnx-vits-rasa-13-exporter. See its README for setup and CLI options.

Licenses and attribution

• Model weights and architecture: CC-BY-4.0. See: ai4bharat/vits_rasa_13.
• ONNX export tooling: MIT. See: github.com/matiaslin/sherpa-onnx-vits-rasa-13-exporter

If you use this model, please cite:

@article{ai4bharat_vits_rasa_13,
  title={VITS TTS for Indian Languages},
  author={Ashwin Sankar},
  year={2024},
  publisher={Hugging Face}
}

See upstream Citation.

Known differences from upstream HF inference

• Duration and prosody may differ from the HF Python reference. sherpa-onnx defaults to length_scale=1.0, while HF's inference uses length_scale=1.0/config.speaking_rate internally. Set --vits-length-scale if you need to match a specific HF invocation.
• sherpa-onnx wraps token ids with an extra BOS/EOS blank (0) around HF's add_blank output (which might result in a few silent frames' difference at each end).