合成语音末尾容易吞音

[cpken]

反复测试后得出结论，不管句子末尾是否有无标点符号，末尾都容易吞音。

合成文本

hello world.

合成结果，文件名代表着对应的音色。

expr-voice-2-m.wav

expr-voice-4-m.wav

expr-voice-5-f.wav

[cpken]

解决方法

kittentts/onnx_model.py

def generate(self, text: str, voice: str = "expr-voice-5-m", speed: float = 1.0) -> np.ndarray:
    ···
    # Trim audio
-   audio = outputs[0][5000:-10000]
+   audio = outputs[0][20:-200]

expr-voice-2-m-改善后.wav

[akx]

My #22 will improve that trim behavior.

[Surealist123]

Thankyou Really appreciated all the help lately... Still trying to work out where I can make any money though!!!, lol

…

On Tue, 19 Aug 2025 at 23:20, Aarni Koskela ***@***.***> wrote: *akx* left a comment (KittenML/KittenTTS#72) <#72 (comment)> My #22 <#22> will improve that trim behavior. — Reply to this email directly, view it on GitHub <#72 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/BSCLPT6CE3ANKU2Z3WWZEV33OMQC3AVCNFSM6AAAAACEDZOBPOVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTEMBQG42DKNBWGQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[andkirby]

The easiest way to avoid it - add '...' in the end.
I added fade-out (to avoid clicks) and this trick in #78 for TTS streaming from CLI. Works good.