Trim generated audio based on edge silence

[akx]

This fixes very short generations ("Hello, world!") being abruptly cut off. The threshold 0.01 was chosen empirically.

[SamueleTorregrossa]

Really nice improvement!

[horatio-sans-serif]

This works a bit better IMHO. I am swamped at work so it's kinda raw atm:

diff --git a/kittentts/onnx_model.py b/kittentts/onnx_model.py
index abc123..def456 100644
--- a/kittentts/onnx_model.py
+++ b/kittentts/onnx_model.py
@@ -100,11 +100,13 @@ class KittenTTS_1_Onnx:
         onnx_inputs = self._prepare_inputs(text, voice, speed)
         
         outputs = self.session.run(None, onnx_inputs)
+        audio = outputs[0]
         
-        # Trim edge silence from audio
-        non_silent = np.abs(audio) >= 0.01
-        if np.any(non_silent):
-            indices = np.where(non_silent)[0]
-            start, end = indices[0], indices[-1]
-            audio = audio[start : end + 1]
+        # Conservative trimming to avoid cutting off the end syllable
+        # The original -10000 was too aggressive and cut off final syllables
+        # This approach keeps more of the ending while still removing most silence
+        audio = audio[5000:-5000]  # Reduced end trim from -10000 to -5000
+        
+        # Add a small amount of padding to ensure no cutoff
+        # This adds ~83ms of silence at 24kHz sample rate
+        audio = np.pad(audio, (0, 2000), 'constant')
 
         return audio

[akx]

This works a bit better IMHO. I am swamped at work so it's kinda raw atm:

It'll still always cut off 5000/24000 samples ~= 0.2 seconds from each end, no matter if there happens to be information there? 🤔

[Mic92]

Noticed that with the latest model this seems no longer needed.

[connecteev]

@Mic92 what is the latest model and how can I get it?

[Mic92]

Copied from the README:

m = KittenTTS("KittenML/kitten-tts-nano-0.2")

Going to huggingface also should work.

[Krystal5222]

Unsubscribe

…

On Thu., Sep. 4, 2025, 9:04 a.m. Jörg Thalheim ***@***.***> wrote: *Mic92* left a comment (KittenML/KittenTTS#22) <#22 (comment)> Copied from the README: m = KittenTTS("KittenML/kitten-tts-nano-0.2") Going to huggingface also should work. — Reply to this email directly, view it on GitHub <#22 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/A5AYBKENOBP6LLTUXILYPQ33RA2EHAVCNFSM6AAAAACDHQ23JWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTENJTGYYTEMZWGQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>