feat: add Chatterbox TTS engine for multilingual voice cloning

- New ChatterboxTTSBackend wrapping ChatterboxMultilingualTTS (ResembleAI/chatterbox)
- Supports 23 languages including Hebrew, forces CPU on macOS (MPS issue)
- Monkey-patches torch.load for CPU loading, forces eager attention for compatibility
- trim_tts_output utility cuts trailing silence/hallucination from Chatterbox output
- Full engine integration: /generate, /generate/stream, model status/download/delete
- Hebrew (he) added to supported languages in frontend and backend validation
- Single flat model dropdown extended with Chatterbox option in both generation UIs
- ModelManagement UI groups LuxTTS and Chatterbox under 'Other Voice Models' section

backend/utils/audio.py

@@ -80,6 +80,95 @@ def save_audio(
     sf.write(path, audio, sample_rate)
 
 
+def trim_tts_output(
+    audio: np.ndarray,
+    sample_rate: int = 24000,
+    frame_ms: int = 20,
+    silence_threshold_db: float = -40.0,
+    min_silence_ms: int = 200,
+    max_internal_silence_ms: int = 1000,
+    fade_ms: int = 30,
+) -> np.ndarray:
+    """
+    Trim trailing silence and post-silence hallucination from TTS output.
+
+    Chatterbox sometimes produces ``[speech][silence][hallucinated noise]``.
+    This detects internal silence gaps longer than *max_internal_silence_ms*
+    and cuts the audio at that boundary, then trims trailing silence and
+    applies a short cosine fade-out.
+
+    Args:
+        audio: Input audio array (mono float32)
+        sample_rate: Sample rate in Hz
+        frame_ms: Frame size for RMS energy calculation
+        silence_threshold_db: dB threshold below which a frame is silence
+        min_silence_ms: Minimum trailing silence to keep
+        max_internal_silence_ms: Cut after any silence gap longer than this
+        fade_ms: Cosine fade-out duration in ms
+
+    Returns:
+        Trimmed audio array
+    """
+    frame_len = int(sample_rate * frame_ms / 1000)
+    if frame_len == 0 or len(audio) < frame_len:
+        return audio
+
+    n_frames = len(audio) // frame_len
+    threshold_linear = 10 ** (silence_threshold_db / 20)
+
+    # Compute per-frame RMS
+    rms = np.array(
+        [
+            np.sqrt(np.mean(audio[i * frame_len : (i + 1) * frame_len] ** 2))
+            for i in range(n_frames)
+        ]
+    )
+    is_speech = rms >= threshold_linear
+
+    # Find first speech frame
+    first_speech = 0
+    for i, s in enumerate(is_speech):
+        if s:
+            first_speech = max(0, i - 1)  # keep 1 frame padding
+            break
+
+    # Walk forward from first speech; cut at long internal silence gaps
+    max_silence_frames = int(max_internal_silence_ms / frame_ms)
+    consecutive_silence = 0
+    cut_frame = n_frames
+
+    for i in range(first_speech, n_frames):
+        if is_speech[i]:
+            consecutive_silence = 0
+        else:
+            consecutive_silence += 1
+            if consecutive_silence >= max_silence_frames:
+                cut_frame = i - consecutive_silence + 1
+                break
+
+    # Trim trailing silence from the cut point
+    min_silence_frames = int(min_silence_ms / frame_ms)
+    end_frame = cut_frame
+    while end_frame > first_speech and not is_speech[end_frame - 1]:
+        end_frame -= 1
+    # Keep a short tail
+    end_frame = min(end_frame + min_silence_frames, cut_frame)
+
+    # Convert frames back to samples
+    start_sample = first_speech * frame_len
+    end_sample = min(end_frame * frame_len, len(audio))
+
+    trimmed = audio[start_sample:end_sample].copy()
+
+    # Cosine fade-out
+    fade_samples = int(sample_rate * fade_ms / 1000)
+    if fade_samples > 0 and len(trimmed) > fade_samples:
+        fade = np.cos(np.linspace(0, np.pi / 2, fade_samples)) ** 2
+        trimmed[-fade_samples:] *= fade
+
+    return trimmed
+
+
 def validate_reference_audio(
     audio_path: str,
     min_duration: float = 2.0,