pocket-tts-tamil-onnx-int4

Pocket TTS Tamil — int4 ONNX, runs in a browser

Text to speech for Tamil that runs entirely on a CPU — in a web page, with no server, no GPU and no API key. The whole pack is 106 MB.

Made by Prasad Vittaldev. If this is useful to you, follow me on LinkedIn — I post each new model as it lands: https://in.linkedin.com/in/prasadvittaldev

Try it in your browser right now: https://prasadtts.vercel.app

Why this exists

Kyutai's Pocket TTS is a genuinely small, genuinely good TTS model — a 6-layer flow-matching language model (89.4M parameters) over the Mimi neural codec at 24 kHz. The weights are published openly under CC-BY-4.0.

Getting it into a browser is a separate problem from getting it small. This pack solves the parts that are not obvious:

• int4, in one file. The language model is quantized to 4-bit MatMulNBits, 302.7 MB → 40.8 MB. Crucially it is saved as a single .onnx — ONNX Runtime Web cannot read an external .onnx.data companion and fails at session creation with Module.MountedFiles is not available. Most ONNX exports of models this size split by default.
• The text conditioner stays float32. Quantizing it to int8 saves 40 bytes (it is nearly all embedding lookups) while putting pronunciation at risk.
• Precomputed voice state. A voice prompt is a KV-cache prefix. Shipping it precomputed means the browser never has to load the Mimi encoder at all. One caveat learned the hard way: that state carries a per-layer step that must equal the cache length. Set it to zero and the model silently ignores the voice and emits near-silence — measured at rms 0.0026 versus 0.176 correct.

Contents

Plus bundle.json (graph/state manifest), tokenizer.model (SentencePiece), and bos_before_voice.npy. Voice embeddings are under voices/.

Voices and samples

Every voice below is a real audio file in this repo — click and listen.

• sample_geetha.wav

Use it

In a browser, the easiest path is https://prasadtts.vercel.app — pick Tamil, and the model downloads once and is cached locally.

With ONNX Runtime in Python:

pip install onnxruntime numpy huggingface_hub soundfile
huggingface-cli download prasadvittaldev/pocket-tts-tamil-onnx-int4 --local-dir ./tamil

Then run it with the reference ONNX runtime from KevinAHM/pocket-tts-onnx-export (MIT), pointing --models_dir at the folder you just downloaded and --precision int4.

Credits

• Kyutai — the Pocket TTS model and the Mimi codec. All the speech quality here is theirs; this pack changes only the numeric format. Licensed CC-BY-4.0, and this derivative keeps that licence.
• KevinAHM/pocket-tts-onnx-export (MIT) — the ONNX export scripts these graphs were produced with.
• Quantization, single-file packing, browser runtime and the web app by Prasad Vittaldev — https://in.linkedin.com/in/prasadvittaldev

If you build something with this, I would like to hear about it.