Modelpublic

OpenVoiceOS/phoonnx-spark-tts

sourceHugging Facecc-by-nc-sa-4.0updated 2mo agoView on Hugging Face

0likes7downloads

Model Card

Spark-TTS 0.5B — ONNX for phoonnx

ONNX weights for Spark-TTS 0.5B (SparkAudio), packaged for the phoonnx sparktts engine.

Spark-TTS is a decoder-only language model on a Qwen2.5-0.5B backbone. It does not predict audio. It predicts BiCodec tokens, and BiCodec turns those tokens into a waveform. BiCodec keeps two streams: 32 global tokens that carry the speaker, and one semantic stream at 50 Hz that carries what is said. The language model reads the text and the 32 global tokens, then writes the semantic stream.

Files

File	Stage	Input	Output
`model.onnx` (+ `model.onnx_data`)	Qwen2 language model, KV-cached	`input_ids`, `attention_mask`, `position_ids`, `past_key_values.*`	`logits`, `present.*`
`bicodec_vocoder.onnx`	BiCodec decoder	`semantic_tokens` `[1,T]`, `global_tokens` `[1,1,32]`	waveform `[1,1,N]` at 16 kHz
`wav2vec2_model.onnx`	cloning front end	`wav` `[1,N]` at 16 kHz, zero-mean unit-variance	`feat` `[1,T,1024]`
`bicodec_encoder_quantizer.onnx`	cloning: semantic stream	`feat` `[1,T,1024]`	`semantic_tokens` `[1,T]`
`speaker_encoder_tokenizer.onnx`	cloning: speaker stream	`spec` `[1,513,T]` magnitude STFT	`global_tokens` `[1,1,32]`
`tokenizer.json`	the model's own subword BPE
`voices/*.json`	preset speakers: 32 global tokens each

The short-time Fourier transform in front of the speaker encoder is not in the graphs. ONNX has no complex dtype, so neither torch exporter can lower torch.stft. The mel filterbank projection is inside speaker_encoder_tokenizer.onnx; a caller only has to supply the magnitude spectrogram of torchaudio.transforms.Spectrogram(n_fft=1024, win_length=640, hop_length=320, power=1).

Provenance

—model.onnx is mirrored from Fhrozen/Spark-TTS-0.5B-ONNX, verified against the SparkAudio torch weights (below).
—the four BiCodec / wav2vec2 graphs were exported from SparkAudio/Spark-TTS-0.5B with opset 17.
—the preset voices were minted with Spark-TTS controllable-TTS mode (gender, moderate pitch, moderate speed) and then frozen, so each voice is stable across calls.

Verified against the torch model

Stage	Metric	Result
Language model, prefill	max absolute logit difference	1.3e-3
Language model, 8 decode steps	max absolute logit difference	6.6e-4
Language model, 8 decode steps	greedy token agreement	8/8
Speaker tokenizer	token mismatches	0/32
Encoder + quantizer	token mismatches	0/149
BiCodec vocoder	max absolute waveform difference	1.2e-6
Magnitude STFT (NumPy vs torchaudio)	max absolute difference	4.8e-7
wav2vec2 features (real clip)	max relative difference	1.8e-5
wav2vec2 → quantizer, end to end	semantic token mismatches	0/199

The quantized language models published upstream (model_q4, model_q4f16, model_int8) are not mirrored here. They diverge from the torch model by tens of logits and disagree with greedy decoding, so they are not safe defaults.

License and attribution

Spark-TTS 0.5B is released by SparkAudio under CC-BY-NC-SA-4.0; the upstream code is Apache-2.0. This repository redistributes converted weights under the same terms and adds no license of its own. Cite the original work:

Spark-TTS: An Efficient LLM-Based Text-to-Speech Model with Single-Stream Decoupled Speech Tokens — SparkAudio. https://github.com/SparkAudio/Spark-TTS