Modelpublic

PulpCut/Qwen3-TTS-12Hz-0.6B-Base-safetensors

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes

Model Card

Qwen3-TTS 12Hz 0.6B Base, repackaged

Qwen3-TTS 12Hz 0.6B Base split into one safetensors file per subsystem, so a C engine can map each on its own and validate them separately. The weights are not retrained, merged, pruned or quantized. If you want the model itself, use the original; this exists so H3ddle can load it without a Python runtime.

What is here

file	what it holds
`talker.safetensors`	28 layers at 1024 wide, plus the 151936 × 2048 text embedding
`code_predictor.safetensors`	the 5-layer model that emits code groups 1–15
`speaker_encoder.safetensors`	ECAPA-TDNN over a 128-band mel
`codec_decoder.safetensors`	RVQ, an 8-layer transformer, and the vocoder
`tokenizer.json`	the vocabulary, byte for byte the released one

What changed, and only this

The release is two checkpoints holding five subsystems between them. Four are needed to speak. Two transforms alter any values, and both are exact or a division the reference performs anyway:

—The codec's codebooks were folded from their EMA form — embedding_sum / max(cluster_usage, 1e-5) — which is the division the reference does on every decode. Reading embedding_sum raw and skipping this yields noise that sounds like a subtle decoder bug.
—The speaker encoder was widened from bfloat16 to float32, which is exact. It is 10M parameters, so the size costs nothing, and it removes a precision question from a net that pools statistics over a whole clip.

Everything else is copied byte for byte. The talker keeps its released bfloat16 and the codec decoder its released float32.

What was left out

encoder.* from the speech tokenizer — a Mimi codec encoder that turns a reference clip into codes for in-context cloning. H3ddle clones from the ECAPA speaker embedding instead, so the encoder is 31 unused quantizers and a convolution stack that would otherwise have to be ported and shipped.

If you need in-context cloning, this package cannot do it. Use the original.

Measured H3ddle performance versus the regular CPU path

These are runtime gains from H3ddle's native Metal execution of the same package; the checkpoint itself is not a speed-distilled model.

comparison	regular / baseline	optimized path	resulting gain
talker + code predictor, six-second utterance	8.48 s CPU	3.66 s Metal	56.8% (2.32x)
codec decode, 79 frames	8.02 s CPU	1.22 s Metal	84.8% (6.57x)
complete 79-frame utterance after codec move	13.20 s	4.90 s	62.9% (2.69x)

The CPU and Metal codec outputs agree within normal F32 accumulation error; both speech paths were listened to and judged correct. Actual time varies with utterance length, sampling, reference audio, and Mac.

Licence

Apache 2.0, from the original. NOTICE records the attribution and the modifications above; LICENSE is the licence text.

Qwen3-TTS is by Alibaba Cloud.