fix(kokoro): trim trailing silence and run-on noise on short prompt synthesis (#960)

backend/backends/__init__.py

@@ -369,6 +369,7 @@ def _get_non_qwen_tts_configs() -> list[ModelConfig]:
             engine="kokoro",
             hf_repo_id="hexgrad/Kokoro-82M",
             size_mb=350,
+            needs_trim=True,
             languages=["en", "es", "fr", "hi", "it", "pt", "ja", "zh"],
         ),
     ]

backend/backends/kokoro_backend.py

@@ -288,7 +288,10 @@ class KokoroTTSBackend:
                 # Return 1 second of silence as fallback
                 return np.zeros(KOKORO_SAMPLE_RATE, dtype=np.float32), KOKORO_SAMPLE_RATE
 
-            audio = np.concatenate(audio_chunks)
+            audio = np.concatenate(audio_chunks).astype(np.float32)
-            return audio.astype(np.float32), KOKORO_SAMPLE_RATE
+            from ..utils.audio import trim_tts_output
+
+            audio = trim_tts_output(audio, sample_rate=KOKORO_SAMPLE_RATE)
+            return audio, KOKORO_SAMPLE_RATE
 
         return await asyncio.to_thread(_generate_sync)

backend/tests/test_kokoro_trim.py

@@ -0,0 +1,66 @@
+"""Test Kokoro short prompt audio trimming and engine config."""
+
+import numpy as np
+import pytest
+
+from backend.backends import engine_needs_trim, get_model_config
+from backend.utils.audio import trim_tts_output
+
+
+def test_kokoro_engine_needs_trim_enabled():
+    """Verify Kokoro engine is registered with needs_trim=True in model config."""
+    assert engine_needs_trim("kokoro") is True
+    config = get_model_config("kokoro")
+    assert config is not None
+    assert config.needs_trim is True
+
+
+def test_kokoro_trim_tts_output_removes_trailing_dead_space():
+    """Verify trim_tts_output removes trailing silence past speech."""
+    sr = 24000
+    speech = np.full(int(sr * 1.5), 0.2, dtype=np.float32)  # 1.5s speech
+    trailing_silence = np.zeros(int(sr * 1.0), dtype=np.float32)  # 1.0s trailing dead space
+    raw_audio = np.concatenate([speech, trailing_silence])
+
+    trimmed = trim_tts_output(raw_audio, sample_rate=sr)
+
+    # Trimming cuts trailing silence from 2.5s down to speech duration boundary (1.5s)
+    expected_dur_samples = int(sr * 1.5)
+    assert len(trimmed) == expected_dur_samples
+    assert len(trimmed) < len(raw_audio)
+
+
+@pytest.mark.asyncio
+async def test_kokoro_backend_generate_applies_trimming(monkeypatch):
+    """Verify KokoroTTSBackend.generate applies trimming on synthesized output."""
+    from backend.backends.kokoro_backend import KokoroTTSBackend, KOKORO_SAMPLE_RATE
+
+    backend = KokoroTTSBackend()
+
+    # Mock _load_model_sync to avoid requiring real model load in pure unit test
+    monkeypatch.setattr(backend, "_load_model_sync", lambda: None)
+    monkeypatch.setattr(backend, "_model", object())
+
+    # Mock KPipeline output to yield audio with 1s trailing silence
+    sr = KOKORO_SAMPLE_RATE
+    speech = np.full(int(sr * 1.0), 0.2, dtype=np.float32)
+    silence = np.zeros(int(sr * 1.0), dtype=np.float32)
+    fake_audio = np.concatenate([speech, silence])
+
+    class FakeResult:
+        def __init__(self, audio):
+            self.audio = audio
+
+    class FakePipeline:
+        def __call__(self, text, voice, speed=1.0):
+            yield FakeResult(fake_audio)
+
+    monkeypatch.setattr(backend, "_get_pipeline", lambda lang: FakePipeline())
+
+    audio, sample_rate = await backend.generate("Read it back to me.", voice_prompt={})
+
+    assert sample_rate == sr
+    # Original fake audio was 2.0s (1s speech + 1s silence).
+    # Trimmed cuts trailing silence down to 1.0s speech boundary.
+    assert len(audio) / sr == pytest.approx(1.0, abs=0.05)
+    assert len(audio) < len(fake_audio)