合成音频只有"哦"的内容

[zyb8543d]

使用示例代码：
python test_tts.py --model ../../../model/omnivoice-q8_0.gguf --text "Hello world" --output out.wav

[lovemefan]

我在linux测试了一下，音频是正常的。请问是在windows平台测试的吗

python python-api-examples/tts/tts-offline.py --model omnivoice-q8_0.gguf --text "Hello world"
rapidspeech version: 0.1.0
Loading model: omnivoice-q8_0.gguf
[2026-07-19 23:37:51] [INFO ] Architecture detected: OmniVoice
[2026-07-19 23:37:51] [WARN ] GPU requested but no supported GPU backend could be initialized. Falling back to CPU.
[2026-07-19 23:37:51] [INFO ] Scheduler: CPU-only mode
[2026-07-19 23:37:52] [INFO ] OmniVoice LM: layers=28 embd=1024 heads=16/8 hdim=128 ff=3072 codebooks=8
[2026-07-19 23:37:52] [INFO ] [LLM] DEBUG: n_vocab=151676 n_embd=1024 n_layer=28 arch=qwen3 arch_key=omnivoice-lm
[2026-07-19 23:37:52] [INFO ] [LLM] Hyperparameters: n_embd=1024, n_layer=28, n_head=16, n_vocab=151676
[2026-07-19 23:37:52] [INFO ] [LLM] Loaded 151387 BPE merges
[2026-07-19 23:37:52] [INFO ] [LLM] Vocabulary loaded: 151676 tokens, BOS=128245, EOS=128247, UNK=128244, PAD=0
[2026-07-19 23:37:52] [INFO ] [LLM] Mapped 28 Qwen3 layers
[2026-07-19 23:37:52] [INFO ] L0 weights: wq=0x56512a2610d0(2097152) wk=0x56512a260bc0(1048576) wv=0x56512a261280(1048576) wo=0x56512a260d70(2097152)
[2026-07-19 23:37:52] [INFO ] L0 weights: ffn_gate=0x56512a260500(3145728) ffn_up=0x56512a2606b0(3145728) ffn_down=0x56512a260350(3145728)
[2026-07-19 23:37:52] [INFO ] L0 norms: attn_norm=0x56512a2601a0 q_norm=0x56512a260f20 k_norm=0x56512a260a10 ffn_norm=0x56512a260860
[2026-07-19 23:37:52] [INFO ] OmniVoice: quant_level=1 (type=8)
[2026-07-19 23:37:52] [INFO ] OmniVoice: using combined audio tensors (embd=yes heads=yes combined=no)
[2026-07-19 23:37:52] [INFO ] OmniVoice: mapped audio tensors (heads=8, embeddings=8)
[BPE] Loaded: 151676 vocab, 151387 merges, eos_id=-1 specials=7
[2026-07-19 23:37:52] [INFO ] Codec encoder: pos_conv_w ne=(128,48,768) type=1
[2026-07-19 23:37:52] [INFO ] OmniVoice Codec: RVQ=8x1024 dim=64, DAC=5 blocks upsample=960x
[2026-07-19 23:37:52] [INFO ] Codec: fc2_w=Y fc2_b=Y c1w=Y c1b=Y c2w=Y c2b=Y sfinal_a=Y
[2026-07-19 23:37:52] [INFO ] Codec: blk0 s1=Y ctw=Y ctb=Y ru_ok=3/3 in=1024 out=512
[2026-07-19 23:37:52] [INFO ] Codec: blk1 s1=Y ctw=Y ctb=Y ru_ok=3/3 in=512 out=256
[2026-07-19 23:37:52] [INFO ] Codec: blk2 s1=Y ctw=Y ctb=Y ru_ok=3/3 in=256 out=128
[2026-07-19 23:37:52] [INFO ] Codec: blk3 s1=Y ctw=Y ctb=Y ru_ok=3/3 in=128 out=64
[2026-07-19 23:37:52] [INFO ] Codec: blk4 s1=Y ctw=Y ctb=Y ru_ok=3/3 in=64 out=32
[2026-07-19 23:37:52] [INFO ] Codec encoder: RVQ project_in=8/8 fc=Y dac_enc=Y hubert=Y sem_enc=Y
[2026-07-19 23:37:52] [INFO ] OmniVoice: encoder weights redirected to GGUF originals (GPU-resident)
[2026-07-19 23:37:52] [INFO ] OmniVoice: model loaded successfully (codec=yes)
[2026-07-19 23:37:52] [INFO ] RapidSpeech context core initialized successfully.
Model:    OmniVoice (24000 Hz)
Backend:  CPU
Params:   instruct="male" lang=English seed=42 steps=32
Text:     'Hello world'
[2026-07-19 23:37:52] [INFO ] [BPE-DIAG] encode 'Hello world' -> 2 tokens: [9707,1879]
[2026-07-19 23:37:52] [INFO ] OmniVoice: tokenized 'Hello world' -> 2 tokens
[2026-07-19 23:37:52] [INFO ] OmniVoice: encoding 'Hello world' weight=10.2 speed=0.5640 -> 35 target frames (ref_frames=0 ref='Nice to meet you.')
[2026-07-19 23:37:52] [INFO ] [BPE-DIAG] encode '<|lang_start|>en<|lang_end|><|instruct_start|>male<|instruct_end|>' -> 6 tokens: [151670,268,151671,151672,36476,151673]
[2026-07-19 23:37:52] [INFO ] [BPE-DIAG] encode '<|text_start|>Hello world<|text_end|>' -> 4 tokens: [151674,9707,1879,151675]
[2026-07-19 23:37:52] [INFO ] OmniVoice Prompt: B'=2 K=8 S=45 N1=6 N2=4 Sref=0 Stgt=35
[2026-07-19 23:37:52] [INFO ] style_ids: 151670 268 151671 151672 36476 151673
[2026-07-19 23:37:52] [INFO ] text_ids: 151674 9707 1879 151675
[2026-07-19 23:37:52] [INFO ] MaskGIT schedule: total=280 steps=32 sum=280
[2026-07-19 23:37:52] [INFO ] MaskGIT: T=35 K=8 S=45 V=1025 steps=32
[2026-07-19 23:37:53] [INFO ] Step0 logit stats: cond_rms=85.0879 uncond_rms=85.0113 diff_rms=0.5869 ratio=0.0069
[2026-07-19 23:38:05] [INFO ] MaskGIT: all tokens unmasked at step 32/32, early stop
Output:   33600 samples @ 24000 Hz (1.40s, elapsed=13.47s, RTF=9.624)
Wrote     output.wav

[zyb8543d]

我在Linux ubuntu上测试的,使用omnivoice-q8_0.gguf 模型合成不正常，合成只有“哦哦哦哦...”的内容，使用omnivoice.gguf模型合成是正常的。

[zyb8543d]

@lovemefan 无法使用GPU 是咋回事啊？

 Naming format: 001.wav (using sequential numbers)
   Prefix: (none)
   Zero padding: 3 digits
------------------------------------------------------------
🔄 Loading model: ../../../model/omnivoice-f16.gguf
[2026-08-06 11:49:08] [INFO ] Architecture detected: OmniVoice
[2026-08-06 11:49:08] [WARN ] GPU requested but no supported GPU backend could be initialized. Falling back to CPU.
[2026-08-06 11:49:08] [INFO ] Scheduler: CPU-only mode

[lovemefan]

@lovemefan 无法使用GPU 是咋回事啊？

 Naming format: 001.wav (using sequential numbers)
   Prefix: (none)
   Zero padding: 3 digits
------------------------------------------------------------
🔄 Loading model: ../../../model/omnivoice-f16.gguf
[2026-08-06 11:49:08] [INFO ] Architecture detected: OmniVoice
[2026-08-06 11:49:08] [WARN ] GPU requested but no supported GPU backend could be initialized. Falling back to CPU.
[2026-08-06 11:49:08] [INFO ] Scheduler: CPU-only mode

编译的时候需要加上 -DRS_CUDA=ON -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc 参数 ，编译cuda的时长会比较久