idle-intelligence/pocket-tts-int8

Pocket TTS — INT8 Quantized

INT8 channel-wise quantized version of kyutai/pocket-tts-without-voice-cloning for browser-based TTS inference via WebAssembly.

Try the demo →

Model Details

	Original	INT8
File	tts_b6369a24.safetensors	model.safetensors
Size	225 MB	132 MB
Dtype	BF16	I8 + BF16 scales
Reduction	—	41%

Quantization Format

Per-output-channel INT8 quantization with BF16 scale factors.

Each weight tensor foo is split into two tensors:

• foo (I8) — quantized weight values
• foo_scale (BF16) — one scale factor per output channel

Dequantization: weight_bf16 = weight_i8 * scale_bf16

See quantize_config.json for machine-readable metadata.

Files

File	Size	Description
model.safetensors	132 MB	Quantized model weights
tokenizer.model	58 KB	SentencePiece unigram tokenizer
quantize_config.json	<1 KB	Quantization parameters

Voice embeddings are unchanged — use them from the original repo.

Usage

This model is designed for use with tts-web, a browser-based TTS engine built with Candle (Rust ML framework) and WebAssembly. Dequantization happens at load time in memory.

Dequantization (Python)

import safetensors
from safetensors.numpy import load_file
import numpy as np

tensors = load_file("model.safetensors")

for name in list(tensors.keys()):
    if name.endswith("_scale"):
        continue
    scale_name = f"{name}_scale"
    if scale_name in tensors:
        weight_i8 = tensors[name].astype(np.float32)
        scale = tensors[scale_name].astype(np.float32)
        tensors[name] = (weight_i8 * scale).astype(np.float16)
        del tensors[scale_name]

Acknowledgments

Based on Kyutai's Pocket TTS — a 100M parameter text-to-speech model.

Disclaimer

This is an independent port by idle intelligence, not affiliated with or endorsed by Kyutai Labs.

License

CC-BY-4.0 (same as the original model).