svara-tts-v1-fp8

Svara TTS v1 (FP8 Quantized)

FP8-quantized version of Svara TTS for lower VRAM usage (~6-8GB).

Original Developer: Kenpath
Base Model: kenpath/svara-tts-v1

Features

• 12 Indian Languages: Hindi, English, Telugu, Tamil, Marathi, Bengali, Kannada, Malayalam, Gujarati, Punjabi, Odia, Assamese
• Male/Female Voices: Support for both genders
• Low VRAM: ~6-8GB memory usage (FP8)
• OpenAI-Compatible: /v1/audio/speech endpoint

Supported Languages

Voice Format

Voice IDs: {language_code}_{gender}

Examples: hi_male, ta_female, en_male, bn_female

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "srikantp/svara-tts-v1-fp8",
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("srikantp/svara-tts-v1-fp8")

Requirements

• NVIDIA GPU with 8GB+ VRAM
• CUDA drivers
• transformers, torch

Inference Code

For inference code, server, and web UI, see: svara-tts-inference-fp8

License

Apache 2.0 - See LICENSE file.

Citation

@misc{svara-tts-v1-fp8,
  title={Svara TTS v1 FP8},
  author={srikantp},
  year={2025},
  url={https://huggingface.co/srikantp/svara-tts-v1-fp8}
}