[BUG] <title>[BUG] TTS audio generation fails with AttributeError in _extract_speech_token

[weklyy]

是否已有关于该错误的issue或讨论？ | Is there an existing issue / discussion for this?

• 我已经搜索过已有的issues和讨论 | I have searched the existing issues / discussions

该问题是否在FAQ中有解答？ | Is there an existing answer for this in FAQ?

• 我已经搜索过FAQ | I have searched FAQ

当前行为 | Current Behavior

MiniCPM-o-4_5 TTS 功能异常报告

环境信息

• 平台: Tencent Cloud Studio (Ubuntu)
• GPU: NVIDIA A10 (24GB)
• CUDA: 12.8
• Python: 3.11.1
• torch: 2.8.0+cu128
• torchaudio: 2.8.0+cu128
• transformers: 4.51.0
• onnxruntime-gpu: 1.24.1
• minicpmo-utils: 1.0.4

问题描述

TTS 语音合成功能无法正常生成音频文件。文本生成正常，但音频文件始终不生成。

错误信息

File "cosyvoice/cli/frontend.py", line 93, in _extract_speech_token assert speech.shape[1] / 16000 <= 30, 'do not support extract speech token for audio longer than 30s' AttributeError: 'NoneType' object has no attribute 'shape'

复现步骤

1. 加载模型: init_tts=True
2. 调用 model.init_tts(streaming=False)
3. 使用 model.chat() 并设置 generate_audio=True
4. 音频文件不生成，报错如上

其他功能测试结果

• ✅ 纯文本对话 - 正常
• ✅ 图片理解 - 正常
• ✅ 语音识别 (ASR) - 正常
• ✅ 视频理解 - 正常
• ❌ 语音合成 (TTS) - 失败

附件

• diagnostic_report.json

diagnostic_report.json

期望行为 | Expected Behavior

No response

复现方法 | Steps To Reproduce

复现步骤

1. 加载模型: init_tts=True
2. 调用 model.init_tts(streaming=False)
3. 使用 model.chat() 并设置 generate_audio=True
4. 音频文件不生成，报错如上

运行环境 | Environment

- OS:
- Python:
- Transformers:
- PyTorch:
- CUDA (`python -c 'import torch; print(torch.version.cuda)'`):

备注 | Anything else?

No response

[airlsyn]

Could you share your input ? Thanks @weklyy

[weklyy]

"Hi, thanks for checking.

The crash happens during the model.chat call when init_tts=True is set.

My Input:

python

1. Model Initialization

model.init_tts(streaming=False)

2. Input Messages

msgs = [{"role": "user", "content": "Please read this text out loud."}]

3. Model Call

res = model.chat(
msgs=msgs,
tokenizer=tokenizer,
generate_audio=True
)
Error: AttributeError: 'NoneType' object has no attribute 'shape' in _extract_speech_token.

It seems like even simple text input triggers this when audio generation is enabled."
@airlsyn

[airlsyn]

@weklyy
The TTS module requires a reference audio to extract speech tokens for voice synthesis. Without it, we cannot generate audio.

You can provide the reference audio in two ways:

Option 1: Via system prompt

ref_audio_path = "path/to/reference_audio.wav"
sys_msg = model.get_sys_prompt(ref_audio=ref_audio_path, mode="omni", language="en")

msgs = [
    sys_msg,
    {"role": "user", "content": ["Describe this image.", image]}
]

res = model.chat(msgs=msgs, generate_audio=True, output_audio_path="output.wav", use_tts_template=True)

Option 2: Directly in the user message

msgs = [
    {
        "role": "user",
        "content": [
            {"type": "audio_url", "audio_url": {"url": "path/to/reference_audio.wav"}},
            "Please read this text out loud."
        ]
    }
]

res = model.chat(msgs=msgs, generate_audio=True, output_audio_path="output.wav", use_tts_template=True)

Note: For Option 2, you may need to adjust the prompt to get better results, e.g. adding instructions like "Clone the voice in the provided audio prompt." to guide the model to use the reference audio as the target voice.

FYI: A sample reference audio is included in the Huggingface repo (MiniCPM-o-4_5) at assets/HT_ref_audio.wav