[LLM:Feature] Support Qwen3-TTS-12Hz-0.6B-Base

[huangzhengxiang]

Description

Support Qwen3-TTS-12Hz-0.6B-Base

Module

LLM.

Exported model at huangzhengxiang/Qwen3-TTS-0.6B-Base-FP16-MNN and huangzhengxiang/Qwen3-TTS-0.6B-Base-INT8-MNN.

TTS demo at transformers/llm/engine/demo/qwen3_tts_demo.cpp. (Currently only support text to speech. Reference voice is not supported in the demo.)

Type

• Feature

Checklist

• Commit message follows [Module:Type] Description format
• Code compiles without errors
• Tested on relevant platform(s)
• No unrelated format or style changes included

[Rye426]

谢谢你的贡献
当我尝试自己导出4bit 的mnn模型时,出现：
python3 transformers/llm/export/llmexport.py --path Qwen3-TTS-12Hz-0.6B-Base/ --export mnn --quant_bit 4 --quant_block 128 --dst_path Qwen3-TTS-12Hz-0.6B-Base/mnn/

---

Warning: flash-attn is not installed. Will only run the manual PyTorch version. Please install flash-attn for faster inference.

---

/bin/sh: line 1: sox: command not found
WARNING:sox:SoX could not be found!

If you do not have SoX, proceed here:
 - - - http://sox.sourceforge.net/ - - -

If you do (or think that you should) have SoX, double-check your
path variables.

2026-07-02 19:39:27.915645554 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card3": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card3/device/vendor"
2026-07-02 19:39:27.915713016 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card1": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card1/device/vendor"
2026-07-02 19:39:27.915733716 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card4": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card4/device/vendor"
2026-07-02 19:39:27.915750747 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card2": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card2/device/vendor"
2026-07-02 19:39:27.915768727 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card0": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card0/device/vendor"
The module name (originally ) is not a valid Python identifier. Please rename the original module to avoid import issues.
✅ Done load pretrained model /path/Qwen3-TTS-12Hz-0.6B-Base/[ 3.62 s]
✅ Done export qwen3_tts code predictor to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/code_predictor.onnx[ 0.84 s]
✅ Done export qwen3_tts codec embedder to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/codec_embedder.onnx[ 0.03 s]
✅ Done export qwen3_tts talker to ['/path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/talker.onnx', '/path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/code_predictor.onnx', '/home/c
✅ Done convert onnx model to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/talker.mnn[ 1.75 s]
Load LayerNorm data: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2743/2743 [00:00<00:00, 919294.92it/s]
Quant weights: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2743/2743 [00:02<00:00, 1101.03it/s]
✅ Done quant model weight to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/talker.mnn.weight[ 2.67 s]
terminate called after throwing an instance of 'std::out_of_range'
what(): vector::_M_range_check: __n (which is 0) >= this->size() (which is 0)
Aborted

然后我尝试用您给的Qwen3-TTS-0.6B-Base-INT8-MNN：
./build/qwen3_tts_demo Qwen3-TTS-0.6B-Base-INT8-MNN --text "一九九三年二月二十三日上午，四川省安岳县安源乡武村彭家姑嫂五人进城购置衣服" 64 ./tts_output/ chinese --normalize 0.95 推理时间比较长，并且只有电流声
当我的text用”你好，千问“，能够正常输出，但RTF值比较大。

[huangzhengxiang]

自己导出的话需要编译MNNConvert并指定--mnnconvert的路径，预编译的MNN python包暂不支持.

[huangzhengxiang]

谢谢你的贡献 当我尝试自己导出4bit 的mnn模型时,出现： python3 transformers/llm/export/llmexport.py --path Qwen3-TTS-12Hz-0.6B-Base/ --export mnn --quant_bit 4 --quant_block 128 --dst_path Qwen3-TTS-12Hz-0.6B-Base/mnn/

Warning: flash-attn is not installed. Will only run the manual PyTorch version. Please install flash-attn for faster inference.

/bin/sh: line 1: sox: command not found WARNING:sox:SoX could not be found!

If you do not have SoX, proceed here:
 - - - http://sox.sourceforge.net/ - - -

If you do (or think that you should) have SoX, double-check your
path variables.

2026-07-02 19:39:27.915645554 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card3": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card3/device/vendor" 2026-07-02 19:39:27.915713016 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card1": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card1/device/vendor" 2026-07-02 19:39:27.915733716 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card4": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card4/device/vendor" 2026-07-02 19:39:27.915750747 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card2": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card2/device/vendor" 2026-07-02 19:39:27.915768727 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card0": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card0/device/vendor" The module name (originally ) is not a valid Python identifier. Please rename the original module to avoid import issues. ✅ Done load pretrained model /path/Qwen3-TTS-12Hz-0.6B-Base/[ 3.62 s] ✅ Done export qwen3_tts code predictor to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/code_predictor.onnx[ 0.84 s] ✅ Done export qwen3_tts codec embedder to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/codec_embedder.onnx[ 0.03 s] ✅ Done export qwen3_tts talker to ['/path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/talker.onnx', '/path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/code_predictor.onnx', '/home/c ✅ Done convert onnx model to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/talker.mnn[ 1.75 s] Load LayerNorm data: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2743/2743 [00:00<00:00, 919294.92it/s] Quant weights: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2743/2743 [00:02<00:00, 1101.03it/s] ✅ Done quant model weight to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/talker.mnn.weight[ 2.67 s] terminate called after throwing an instance of 'std::out_of_range' what(): vector::_M_range_check: __n (which is 0) >= this->size() (which is 0) Aborted

然后我尝试用您给的Qwen3-TTS-0.6B-Base-INT8-MNN： ./build/qwen3_tts_demo Qwen3-TTS-0.6B-Base-INT8-MNN --text "一九九三年二月二十三日上午，四川省安岳县安源乡武村彭家姑嫂五人进城购置衣服" 64 ./tts_output/ chinese --normalize 0.95 推理时间比较长，并且只有电流声 当我的text用”你好，千问“，能够正常输出，但RTF值比较大。

预导出模型的问题我看看能否复现，来解决一下。

[huangzhengxiang]

谢谢你的贡献 当我尝试自己导出4bit 的mnn模型时,出现： python3 transformers/llm/export/llmexport.py --path Qwen3-TTS-12Hz-0.6B-Base/ --export mnn --quant_bit 4 --quant_block 128 --dst_path Qwen3-TTS-12Hz-0.6B-Base/mnn/

Warning: flash-attn is not installed. Will only run the manual PyTorch version. Please install flash-attn for faster inference.

/bin/sh: line 1: sox: command not found WARNING:sox:SoX could not be found!

If you do not have SoX, proceed here:
 - - - http://sox.sourceforge.net/ - - -

If you do (or think that you should) have SoX, double-check your
path variables.

2026-07-02 19:39:27.915645554 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card3": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card3/device/vendor" 2026-07-02 19:39:27.915713016 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card1": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card1/device/vendor" 2026-07-02 19:39:27.915733716 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card4": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card4/device/vendor" 2026-07-02 19:39:27.915750747 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card2": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card2/device/vendor" 2026-07-02 19:39:27.915768727 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card0": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card0/device/vendor" The module name (originally ) is not a valid Python identifier. Please rename the original module to avoid import issues. ✅ Done load pretrained model /path/Qwen3-TTS-12Hz-0.6B-Base/[ 3.62 s] ✅ Done export qwen3_tts code predictor to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/code_predictor.onnx[ 0.84 s] ✅ Done export qwen3_tts codec embedder to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/codec_embedder.onnx[ 0.03 s] ✅ Done export qwen3_tts talker to ['/path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/talker.onnx', '/path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/code_predictor.onnx', '/home/c ✅ Done convert onnx model to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/talker.mnn[ 1.75 s] Load LayerNorm data: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2743/2743 [00:00<00:00, 919294.92it/s] Quant weights: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2743/2743 [00:02<00:00, 1101.03it/s] ✅ Done quant model weight to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/talker.mnn.weight[ 2.67 s] terminate called after throwing an instance of 'std::out_of_range' what(): vector::_M_range_check: __n (which is 0) >= this->size() (which is 0) Aborted

然后我尝试用您给的Qwen3-TTS-0.6B-Base-INT8-MNN： ./build/qwen3_tts_demo Qwen3-TTS-0.6B-Base-INT8-MNN --text "一九九三年二月二十三日上午，四川省安岳县安源乡武村彭家姑嫂五人进城购置衣服" 64 ./tts_output/ chinese --normalize 0.95 推理时间比较长，并且只有电流声 当我的text用”你好，千问“，能够正常输出，但RTF值比较大。

问题已修复，是没有加reference voice参考音频的原因。官方Demo需要添加一段声音作为参考音频才能正常生成。
现在您尝试
build-linux/qwen3_tts_demo
model/Qwen3-TTS-0.6B-Base-MNN-int8
--text "一九九三年二月二十三日上午，四川省安岳县安源乡武村彭家姑嫂五人进城购置衣服"
64
./tts_output/
chinese
--ref_audio apps/iOS/MNNLLMChat/MNNLLMiOS/LocalBatchTest/Audios/BAC009S0764W0121.wav
--normalize 0.95

生成音频在：
./tts_output/qwen3_tts_text.wav

如果要对比 fp16，把模型路径替换成：
model/Qwen3-TTS-0.6B-Base-MNN-fp16

[Rye426]

谢谢你的贡献 当我尝试自己导出4bit 的mnn模型时,出现： python3 transformers/llm/export/llmexport.py --path Qwen3-TTS-12Hz-0.6B-Base/ --export mnn --quant_bit 4 --quant_block 128 --dst_path Qwen3-TTS-12Hz-0.6B-Base/mnn/
Warning: flash-attn is not installed. Will only run the manual PyTorch version. Please install flash-attn for faster inference.
/bin/sh: line 1: sox: command not found WARNING:sox:SoX could not be found!

If you do not have SoX, proceed here:
 - - - http://sox.sourceforge.net/ - - -

If you do (or think that you should) have SoX, double-check your
path variables.

2026-07-02 19:39:27.915645554 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card3": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card3/device/vendor" 2026-07-02 19:39:27.915713016 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card1": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card1/device/vendor" 2026-07-02 19:39:27.915733716 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card4": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card4/device/vendor" 2026-07-02 19:39:27.915750747 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card2": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card2/device/vendor" 2026-07-02 19:39:27.915768727 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card0": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card0/device/vendor" The module name (originally ) is not a valid Python identifier. Please rename the original module to avoid import issues. ✅ Done load pretrained model /path/Qwen3-TTS-12Hz-0.6B-Base/[ 3.62 s] ✅ Done export qwen3_tts code predictor to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/code_predictor.onnx[ 0.84 s] ✅ Done export qwen3_tts codec embedder to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/codec_embedder.onnx[ 0.03 s] ✅ Done export qwen3_tts talker to ['/path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/talker.onnx', '/path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/code_predictor.onnx', '/home/c ✅ Done convert onnx model to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/talker.mnn[ 1.75 s] Load LayerNorm data: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2743/2743 [00:00<00:00, 919294.92it/s] Quant weights: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2743/2743 [00:02<00:00, 1101.03it/s] ✅ Done quant model weight to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/talker.mnn.weight[ 2.67 s] terminate called after throwing an instance of 'std::out_of_range' what(): vector::_M_range_check: __n (which is 0) >= this->size() (which is 0) Aborted
然后我尝试用您给的Qwen3-TTS-0.6B-Base-INT8-MNN： ./build/qwen3_tts_demo Qwen3-TTS-0.6B-Base-INT8-MNN --text "一九九三年二月二十三日上午，四川省安岳县安源乡武村彭家姑嫂五人进城购置衣服" 64 ./tts_output/ chinese --normalize 0.95 推理时间比较长，并且只有电流声 当我的text用”你好，千问“，能够正常输出，但RTF值比较大。

问题已修复，是没有加reference voice参考音频的原因。官方Demo需要添加一段声音作为参考音频才能正常生成。 现在您尝试 build-linux/qwen3_tts_demo model/Qwen3-TTS-0.6B-Base-MNN-int8 --text "一九九三年二月二十三日上午，四川省安岳县安源乡武村彭家姑嫂五人进城购置衣服" 64 ./tts_output/ chinese --ref_audio apps/iOS/MNNLLMChat/MNNLLMiOS/LocalBatchTest/Audios/BAC009S0764W0121.wav --normalize 0.95

生成音频在： ./tts_output/qwen3_tts_text.wav

如果要对比 fp16，把模型路径替换成： model/Qwen3-TTS-0.6B-Base-MNN-fp16

根据以上命令，这是以下输出
CPU Group: [ 1 2 3 4 ], 799999 - 1799998
CPU Group: [ 7 8 ], 799830 - 2199850
CPU Group: [ 5 6 ], 799830 - 2299829
CPU Group: [ 9 10 ], 799830 - 2499787
CPU Group: [ 0 11 ], 799830 - 2600083
The device supports: i8sdot:1, fp16:1, i8mm: 1, sve2: 1, sme2: 0
waveform shape=[1,122880] size=122880
waveform first16: 0.00023078975 0.0002876065 0.00030268927 0.00032124258 0.00024789077 2.1432717e-05 -0.00028724791 -7.7549652e-05 2.0121633e-05 -3.7976828e-05 -0.00016274943 -0.00020869241 -0.00020560113 -0.00023263895 -0.00026221521 -0.00029193162
normalize target_peak=0.95 original_peak=1 scale=0.95
waveform_normalized shape=[1,122880] size=122880
waveform_normalized first16: 0.00021925026 0.00027322618 0.00028755481 0.00030518044 0.00023549623 2.0361082e-05 -0.00027288552 -7.367217e-05 1.9115552e-05 -3.6077985e-05 -0.00015461196 -0.00019825778 -0.00019532107 -0.00022100699 -0.00024910446 -0.00027733503
saved wav: /home/cix/workspace/TTS_llama.cpp/tts_output/qwen3_tts_text.wav
Qwen3-TTS text C++ chain finished. frames=64

生成的语音不完整只有一半，把64提高到128就完整了，目前生成的时间貌似比较慢，另外可以添加一个计算rtf的功能吗？

[huangzhengxiang]

谢谢你的贡献 当我尝试自己导出4bit 的mnn模型时,出现： python3 transformers/llm/export/llmexport.py --path Qwen3-TTS-12Hz-0.6B-Base/ --export mnn --quant_bit 4 --quant_block 128 --dst_path Qwen3-TTS-12Hz-0.6B-Base/mnn/
Warning: flash-attn is not installed. Will only run the manual PyTorch version. Please install flash-attn for faster inference.
/bin/sh: line 1: sox: command not found WARNING:sox:SoX could not be found!

If you do not have SoX, proceed here:
 - - - http://sox.sourceforge.net/ - - -

If you do (or think that you should) have SoX, double-check your
path variables.

2026-07-02 19:39:27.915645554 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card3": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card3/device/vendor" 2026-07-02 19:39:27.915713016 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card1": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card1/device/vendor" 2026-07-02 19:39:27.915733716 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card4": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card4/device/vendor" 2026-07-02 19:39:27.915750747 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card2": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card2/device/vendor" 2026-07-02 19:39:27.915768727 [W:onnxruntime:Default, device_discovery.cc:283 GetGpuDevices] Failed to detect devices under "/sys/class/drm/card0": device_discovery.cc:93 ReadFileContents Failed to open file: "/sys/class/drm/card0/device/vendor" The module name (originally ) is not a valid Python identifier. Please rename the original module to avoid import issues. ✅ Done load pretrained model /path/Qwen3-TTS-12Hz-0.6B-Base/[ 3.62 s] ✅ Done export qwen3_tts code predictor to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/code_predictor.onnx[ 0.84 s] ✅ Done export qwen3_tts codec embedder to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/codec_embedder.onnx[ 0.03 s] ✅ Done export qwen3_tts talker to ['/path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/talker.onnx', '/path/Qwen3-TTS-12Hz-0.6B-Base/mnn/onnx/code_predictor.onnx', '/home/c ✅ Done convert onnx model to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/talker.mnn[ 1.75 s] Load LayerNorm data: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2743/2743 [00:00<00:00, 919294.92it/s] Quant weights: 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 2743/2743 [00:02<00:00, 1101.03it/s] ✅ Done quant model weight to /path/Qwen3-TTS-12Hz-0.6B-Base/mnn/talker.mnn.weight[ 2.67 s] terminate called after throwing an instance of 'std::out_of_range' what(): vector::_M_range_check: __n (which is 0) >= this->size() (which is 0) Aborted
然后我尝试用您给的Qwen3-TTS-0.6B-Base-INT8-MNN： ./build/qwen3_tts_demo Qwen3-TTS-0.6B-Base-INT8-MNN --text "一九九三年二月二十三日上午，四川省安岳县安源乡武村彭家姑嫂五人进城购置衣服" 64 ./tts_output/ chinese --normalize 0.95 推理时间比较长，并且只有电流声 当我的text用”你好，千问“，能够正常输出，但RTF值比较大。

问题已修复，是没有加reference voice参考音频的原因。官方Demo需要添加一段声音作为参考音频才能正常生成。 现在您尝试 build-linux/qwen3_tts_demo model/Qwen3-TTS-0.6B-Base-MNN-int8 --text "一九九三年二月二十三日上午，四川省安岳县安源乡武村彭家姑嫂五人进城购置衣服" 64 ./tts_output/ chinese --ref_audio apps/iOS/MNNLLMChat/MNNLLMiOS/LocalBatchTest/Audios/BAC009S0764W0121.wav --normalize 0.95
生成音频在： ./tts_output/qwen3_tts_text.wav
如果要对比 fp16，把模型路径替换成： model/Qwen3-TTS-0.6B-Base-MNN-fp16

根据以上命令，这是以下输出 CPU Group: [ 1 2 3 4 ], 799999 - 1799998 CPU Group: [ 7 8 ], 799830 - 2199850 CPU Group: [ 5 6 ], 799830 - 2299829 CPU Group: [ 9 10 ], 799830 - 2499787 CPU Group: [ 0 11 ], 799830 - 2600083 The device supports: i8sdot:1, fp16:1, i8mm: 1, sve2: 1, sme2: 0 waveform shape=[1,122880] size=122880 waveform first16: 0.00023078975 0.0002876065 0.00030268927 0.00032124258 0.00024789077 2.1432717e-05 -0.00028724791 -7.7549652e-05 2.0121633e-05 -3.7976828e-05 -0.00016274943 -0.00020869241 -0.00020560113 -0.00023263895 -0.00026221521 -0.00029193162 normalize target_peak=0.95 original_peak=1 scale=0.95 waveform_normalized shape=[1,122880] size=122880 waveform_normalized first16: 0.00021925026 0.00027322618 0.00028755481 0.00030518044 0.00023549623 2.0361082e-05 -0.00027288552 -7.367217e-05 1.9115552e-05 -3.6077985e-05 -0.00015461196 -0.00019825778 -0.00019532107 -0.00022100699 -0.00024910446 -0.00027733503 saved wav: /home/cix/workspace/TTS_llama.cpp/tts_output/qwen3_tts_text.wav Qwen3-TTS text C++ chain finished. frames=64

生成的语音不完整只有一半，把64提高到128就完整了，目前生成的时间貌似比较慢，另外可以添加一个计算rtf的功能吗？

对的，这个64 128的数字是用来提前截断限制生成长度的，就和max_new_token一样，可以设置大一些，它生成完整会自己终止。
生成时间的话可以在ARM架构手机上跑一下，ARM的CPU会比Intel CPU在计算上快几倍。

[huangzhengxiang]

rtf 就先不加啦，暂时可以把生成的文件保存下来在MNN外面计算一下。