luweigen/qwen3-tts-0.6b-android

Qwen3-TTS-12Hz-0.6B-Base — models prepared for Android

Derived from Qwen/Qwen3-TTS-12Hz-0.6B-Base (Apache-2.0, revision 5d839924) for the native Android read-aloud app in the read-android project (on-device voice-cloning TTS of Markdown / text documents, no network at runtime). Files are downloaded by the app and checked against the sha256 values compiled into the APK. See the original model card for intended use, limitations and the license.

No voice is included. The app clones a voice from a reference recording; the voice pack (speaker x-vector, reference audio codes and reference text, produced on a PC by tools/make-voice.py) is made by the user from a recording they have the right to use and imported into the app separately.

File	What	Runs on
qwen3-tts-0.6b-cp-q4_0.gguf	Code predictor (5-layer Qwen3 that predicts codebooks 1–15) exported as a plain qwen3 GGUF without vocabulary (inputs are embeddings, output is the hidden state), Q4_0	llama.cpp (CPU)
qwen3-tts-0.6b-cp-tables.bin	Code predictor tables, f16: per-step input embedding tables projected to its hidden size, the 15 output heads and the feedback embeddings (the heads are evaluated with NEON in the app)	CPU
qwen3-tts-0.6b-talker-q8_0.gguf	Talker (Qwen3 LM that generates codebook 0 per frame), GGUF qwen3tts architecture converted with llama.cpp b12319 convert_hf_to_gguf.py, then quantised (Q8_0)	llama.cpp (CPU / OpenCL)
rvq-tables.bin	RVQ dequantisation tables, f16 [16][2048][512] (output_proj(codebook_q)), header RTTSRVQ1	CPU
vocoder-w45.c0of3.fp16.tflite	The same three LiteRT segments for the tail window W = 25 + 20 frames (used for the last ≤ 20 frames of a segment; optional)	Adreno GPU, LiteRT 2.3.2 (tested on Adreno 630)
vocoder-w45.c1of3.fp16.tflite	The same three LiteRT segments for the tail window W = 25 + 20 frames (used for the last ≤ 20 frames of a segment; optional)	Adreno GPU, LiteRT 2.3.2 (tested on Adreno 630)
vocoder-w45.c2of3.fp16.tflite	The same three LiteRT segments for the tail window W = 25 + 20 frames (used for the last ≤ 20 frames of a segment; optional)	Adreno GPU, LiteRT 2.3.2 (tested on Adreno 630)
vocoder-w75.c0of3.fp16.tflite	Vocoder W75 as three LiteRT (TFLite) segments with fp16 weights, run one after another on the GPU; segments 0 and 2 need fp32 GPU precision. Mathematically equivalent rewrites for the mobile GPU: per-channel scales folded into neighbouring linear layers, hand-written masked softmax	Adreno GPU, LiteRT 2.3.2 (tested on Adreno 630)
vocoder-w75.c1of3.fp16.tflite	Vocoder W75 as three LiteRT (TFLite) segments with fp16 weights, run one after another on the GPU; segments 0 and 2 need fp32 GPU precision. Mathematically equivalent rewrites for the mobile GPU: per-channel scales folded into neighbouring linear layers, hand-written masked softmax	Adreno GPU, LiteRT 2.3.2 (tested on Adreno 630)
vocoder-w75.c2of3.fp16.tflite	Vocoder W75 as three LiteRT (TFLite) segments with fp16 weights, run one after another on the GPU; segments 0 and 2 need fp32 GPU precision. Mathematically equivalent rewrites for the mobile GPU: per-channel scales folded into neighbouring linear layers, hand-written masked softmax	Adreno GPU, LiteRT 2.3.2 (tested on Adreno 630)
vocoder-w75.htp.onnx	The same vocoder (W75) as a plain fp32 ONNX with static shapes	any device, ONNX Runtime CPU EP (slow)

Pipeline (12.5 frames per second of audio): the talker generates codebook 0 per frame, the code predictor the other 15 codebooks, the vocoder (the speech tokenizer's decoder) turns 16 codebooks × frames into 24 kHz audio in fixed windows of 25 context + N new frames. The decoder is causal, so a window that is not full is padded on the right without changing the result. The RVQ dequantisation is linear and done on the CPU with the precomputed tables in rvq-tables.bin (Σ_q table_q[code_q]); the vocoder files take the latent [1, 512, W].