Spark-TTS-0.5B-ONNX

Spark-TTS-0.5B — ONNX

The CPU half of Spark-TTS-0.5B, as ONNX graphs, so it runs without torch.

spark-speaker-encoder.onnx

Exported here, because it exists in no other public repo: upstream ships it only as a submodule of the 625 MB BiCodec checkpoint, most of which is the decoder. It is BiCodec.speaker_encoder.tokenize — ECAPA-TDNN + perceiver resampler + FSQ — at opset 17, with dynamic batch and frame axes.

Input mel (batch, 128, frames) float32, output speaker_tokens (batch, 1, 32) int64. The mel is 16 kHz, n_fft 1024, win 640, hop 320, f_min 10, 128 mels, power 1, Slaney norm and scale.

Speaker identity depends on the mel alone — verified bit-identical to the full BiCodec.tokenize pipeline, which also loads wav2vec2 (1.2 GB) and the BiCodec encoder (122 MB) to produce semantic tokens this path does not use. Enrollment reads a fixed 6.00 s window and is deterministic.

Produced by scripts/export_spark_speaker_encoder.py in Orb, which verifies the export against torch on a real clip, across frame counts 51/151/301/601, at batch 2, and for repeatability before keeping the file.

sha256 808b0c09187fe031d896c5fdb92d6fc30fd9ac46f951b95988f6065d60ece22c

bicodec.onnx

Byte-identical mirror of Fhrozen/Spark-TTS-0.5B-ONNX, mirrored so a pinned consumer does not depend on a third-party repo staying up. Verified against the torch decoder at 114 dB SNR (correlation 1.00000000), holding from 1 to 1500 semantic tokens.

Inputs semantic_tokens (batch, length) int64 and global_tokens (batch, 1, 32) int64; output audio float32 at 16 kHz.

sha256 a1459483cf562e9de8b9d6077e9246b38df797b81d392b6c5a3cb78a595bb7df

License

CC-BY-NC-SA-4.0, inherited from the base model. Non-commercial.