Reza2kn/gooya-v1-ONNX-int4

Gooya v1 — ONNX Web INT4

Smaller browser-ready INT4 ONNX derivative of Reza2kn/gooya-v1, with the required hybrid MOSS audio codec and the Negara v7 Persian G2P frontend.

New long-form checkpoint

This export was created from the new source revision 9320f582245d8877f207d5609ff3ca1391c7085f using exactly:

• Transformers 4.57.1
• PyTorch 2.12.0
• ONNX opset 17
• all 16 MOSS audio codebooks

The fixed browser sampler bakes the checkpoint card's cooler long-form defaults: audio temperature 0.6, top-p 0.9, top-k 25, and repetition penalty 1.2. Do not recreate it with Transformers 5.x; this remote-code model is verified with Transformers 4.57.1.

Repository layout

MOSS-TTS-Nano-100M-ONNX/       INT4 TTS graphs + shared external weights
MOSS-Audio-Tokenizer-Nano-ONNX/ FP16 full decoder + FP32 encoder/streaming decoder
negara-g2p-v7/onnx/            Persian grapheme-to-phoneme ONNX graphs
validation/                     corrected and broken comparison WAVs

The TTS MatMul weights use asymmetric block-32 MatMulNBits INT4 with FP32 graph I/O. The full codec decoder is FP16; the reference encoder and streaming decoder remain FP32. This is the low-download option. The FP16 repository remains the recommended speed path because WebGPU autoregressive generation measured substantially faster in ONNX Runtime Web 1.27.

Validation

Input: سلام. → s a l A m, fixed sampling, seed 1234, all 16 codebooks.

Build	Frames	Termination	Persian Whisper result
New checkpoint FP16	22	audio EOS	سالو (Whisper-base smoke)
New checkpoint INT4 block-64	15	audio EOS	سالا (Whisper-base smoke)
New checkpoint INT4 block-32	17	audio EOS	سالام (Whisper-base smoke)

Actual browser profiling on the included 68-character Persian demo, ONNX Runtime Web 1.27 WebGPU:

• FP16 synthesis pipeline: 10.86 s before GPU-resident KV-cache optimization
• FP16 synthesis pipeline: 6.22 s with GPU-resident KV cache
• INT4 block-32 synthesis pipeline: 33.28 s; smaller download, but not the speed default
• greedy Negara G2P matched 5-beam output on five representative Persian sentences while reducing warm G2P latency from 6.7–8.1 s to 1.3–1.8 s

The web app overlaps G2P with model loading, caches shared sidecar reads, compiles only the active sampler, keeps autoregressive KV tensors on GPU, and prewarms after the first text interaction.

License

This derivative follows the source model's other license classification. Review the upstream Gooya and OpenMOSS licensing terms before redistribution or commercial use.