Set speaking speed (length_scale): male 1.43, female 1.11

9b48190 verified about 15 hours ago

3.15 kB

language:
  - saz
  - ta
license: cc-by-4.0
tags:
  - text-to-speech
  - tts
  - vits
  - sourashtra
  - low-resource
pipeline_tag: text-to-speech

Sourashtra VITS TTS Models

VITS text-to-speech models for the Sourashtra language (ISO 639-3: saz), a minority Indo-Aryan language spoken primarily in Tamil Nadu, India. Trained using Coqui TTS on a custom annotated speech corpus.

Four variants: 2 speakers (male, female) × 2 input scripts (Tamil script, Sourashtra script).

Models

Folder	Speaker	Input Script	Training Steps
`Sourashtra-Male_Script-tamil`	Male	Tamil (தமிழ்)	300,000
`Sourashtra-Male_Script-sourashtra`	Male	Sourashtra (ꢪꢾꢥꢶꢒ)	300,000
`Sourashtra-Female_Script-tamil`	Female	Tamil (தமிழ்)	340,000
`Sourashtra-Female_Script-sourashtra`	Female	Sourashtra (ꢪꢾꢥꢶꢒ)	340,000

Each folder contains best_model.pth, config.json, inference.py, and requirements.txt.

Setup

pip install -r requirements.txt

For GPU inference, install the CUDA-enabled PyTorch build matching your driver first — see pytorch.org.

Usage

Run inference.py from inside the model folder:

# Male — Tamil script
cd Sourashtra-Male_Script-tamil
python inference.py "சொராஷ்ட்ர மொழி" -o output.wav

# Male — Sourashtra script
cd Sourashtra-Male_Script-sourashtra
python inference.py "ꢪꢾꢥꢶꢒ ꢪꢒꢡ" -o output.wav

# Female — Tamil script
cd Sourashtra-Female_Script-tamil
python inference.py "சொராஷ்ட்ர மொழி" -o output.wav

# Female — Sourashtra script
cd Sourashtra-Female_Script-sourashtra
python inference.py "ꢪꢾꢥꢶꢒ ꢪꢒꢡ" -o output.wav

Use --gpu <id> to select a GPU, or --cpu to force CPU inference.

Script Notes

The Tamil-script and Sourashtra-script models produce speech from the same speaker — only the input orthography differs. Choose based on your text source.

Tamil script models — strip :, ., ' and apply NFC normalization automatically
Sourashtra script models — strip Sourashtra Danda (꣎) and Double Danda (꣏) automatically

Speaking Speed

Each config.json sets model_args.length_scale, which controls speaking speed (higher = slower):

Speaker	`length_scale`
Male	1.43
Female	1.11

At the default of 1.0 the male models speak too fast (about 1/0.7 of natural speed) and the female models slightly fast (about 1/0.9). To change the speed, edit length_scale in the model's config.json.

Training

Parameter	Value
Architecture	VITS (end-to-end, flow-based)
Sample rate	22050 Hz
Mel bins	80
Batch size	16
Mixed precision	Yes
Phonemes	No (character-level)

Male training data: ~9,800–10,000 utterances. Female training data: ~11,400 utterances.