fix(audio): preprocess reference samples instead of rejecting them (#502)

* fix(audio): preprocess reference samples instead of rejecting them

Uploaded/recorded voice samples were rejected outright whenever the peak
exceeded 0.99 ("Audio is clipping (reduce input gain)"). That wasn't
actionable: a recording in the app has no pre-gain control, and an
already-captured file can't be re-taken by the user. The Settings
"Normalize audio" toggle only affects generated TTS output, so users who
enabled it expecting it to help with sample uploads were still blocked.

Replace the hard reject with a small, always-on preprocess step that
runs right after load:
  - DC-offset removal
  - Conservative edge-silence trim (top_db=30) with 100 ms padding kept
  - Peak cap at 0.95 if the input peak exceeds that

Duration and RMS checks now run on the preprocessed waveform, so
samples that were previously rejected for being "hot" are accepted and
stored with safe headroom. True in-waveform clipping artifacts still
can't be repaired — peak scaling only prevents downstream re-clipping
during multi-sample combination and TTS inference.

Adds a unit-test file (previously none existed for audio.py).

Fixes #456.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

* fix(audio): raise trim threshold, cap pad at net-neutral

Review feedback on the preprocessor:

1. ``trim_top_db=30`` was labelled "conservative" in the docstring but is
   actually *more* aggressive than librosa's default of 60. Normal
   speech dynamic range sits around 30 dB, so 30 dB would eat quiet
   trailing syllables and soft consonants. Raise the default to 40 dB —
   below normal speech dynamic range but still catching obvious edge
   silence — and fix the docstring.

2. Unconditional 100 ms edge padding ran even when ``librosa.effects.trim``
   removed nothing. For a well-recorded 29.9 s upload that path would
   push the waveform past the 30 s ceiling and trigger a spurious "too
   long" rejection. Only pad when trimming actually shortened the
   audio, and cap the pad so the output never exceeds the input length.

Adds a regression test for the net-neutral length behaviour.

Co-Authored-By: Claude Opus 4.7 (1M context) <[email protected]>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <[email protected]>

backend/tests/test_audio_preprocess.py

@@ -0,0 +1,112 @@
+"""
+Unit tests for reference-audio preprocessing.
+
+Covers :func:`backend.utils.audio.preprocess_reference_audio` and
+:func:`backend.utils.audio.validate_and_load_reference_audio`.
+"""
+
+import sys
+from pathlib import Path
+
+import numpy as np
+import pytest
+import soundfile as sf
+
+sys.path.insert(0, str(Path(__file__).parent.parent))
+
+from utils.audio import (  # noqa: E402
+    preprocess_reference_audio,
+    validate_and_load_reference_audio,
+)
+
+
+SR = 24000
+
+
+def _tone(duration_s: float, amp: float = 0.3, freq: float = 220.0) -> np.ndarray:
+    n = int(duration_s * SR)
+    t = np.arange(n, dtype=np.float32) / SR
+    return (amp * np.sin(2 * np.pi * freq * t)).astype(np.float32)
+
+
+def test_peak_cap_scales_hot_input():
+    audio = _tone(3.0, amp=0.99)
+    out = preprocess_reference_audio(audio, SR)
+    assert np.abs(out).max() <= 0.951
+
+
+def test_peak_cap_leaves_moderate_input_untouched():
+    audio = _tone(3.0, amp=0.5)
+    out = preprocess_reference_audio(audio, SR)
+    assert np.isclose(np.abs(out).max(), 0.5, atol=1e-3)
+
+
+def test_dc_offset_removed():
+    audio = _tone(3.0, amp=0.3) + 0.1
+    out = preprocess_reference_audio(audio, SR)
+    assert abs(float(np.mean(out))) < 1e-3
+
+
+def test_silence_is_trimmed_with_padding_kept():
+    silence = np.zeros(int(SR * 1.0), dtype=np.float32)
+    speech = _tone(3.0, amp=0.3)
+    audio = np.concatenate([silence, speech, silence])
+    out = preprocess_reference_audio(audio, SR)
+    # Most of the 2s of leading/trailing silence should be gone, but the
+    # 3s of speech plus ~200ms of padding should remain.
+    assert len(audio) - len(out) >= SR, "expected >=1s of silence trimmed"
+    assert len(out) >= int(3.0 * SR), "speech body should be preserved"
+
+
+def test_clean_audio_is_not_padded_past_original_length():
+    # Well-recorded audio with no edge silence shouldn't get longer after
+    # preprocessing — otherwise a 29.9 s upload could be pushed past the
+    # 30 s max_duration ceiling downstream.
+    audio = _tone(3.0, amp=0.3)
+    out = preprocess_reference_audio(audio, SR)
+    assert len(out) <= len(audio)
+
+
+def test_empty_input_returns_empty():
+    out = preprocess_reference_audio(np.zeros(0, dtype=np.float32), SR)
+    assert out.size == 0
+
+
+def test_validate_accepts_previously_rejected_hot_file(tmp_path):
+    audio = _tone(3.0, amp=0.995)
+    path = tmp_path / "hot.wav"
+    sf.write(str(path), audio, SR)
+
+    ok, err, out_audio, out_sr = validate_and_load_reference_audio(str(path))
+
+    assert ok, f"expected pass, got error: {err}"
+    assert out_audio is not None
+    assert out_sr == SR
+    assert np.abs(out_audio).max() <= 0.951
+
+
+def test_validate_still_rejects_silent_input(tmp_path):
+    audio = np.zeros(int(SR * 3.0), dtype=np.float32)
+    path = tmp_path / "silent.wav"
+    sf.write(str(path), audio, SR)
+
+    ok, err, _, _ = validate_and_load_reference_audio(str(path))
+
+    assert not ok
+    assert err is not None
+    assert "too short" in err.lower() or "quiet" in err.lower()
+
+
+def test_validate_rejects_too_short(tmp_path):
+    audio = _tone(0.5, amp=0.3)
+    path = tmp_path / "short.wav"
+    sf.write(str(path), audio, SR)
+
+    ok, err, _, _ = validate_and_load_reference_audio(str(path))
+
+    assert not ok
+    assert "too short" in (err or "").lower()
+
+
+if __name__ == "__main__":
+    pytest.main([__file__, "-v"])

backend/utils/audio.py

@@ -199,6 +199,66 @@ def trim_tts_output(
     return trimmed
 
 
+def preprocess_reference_audio(
+    audio: np.ndarray,
+    sample_rate: int,
+    peak_target: float = 0.95,
+    trim_top_db: float = 40.0,
+    edge_padding_ms: int = 100,
+) -> np.ndarray:
+    """
+    Clean up a reference-audio sample before validation/storage.
+
+    Removes DC offset, trims leading/trailing silence, and caps the peak so a
+    slightly-hot recording doesn't get rejected downstream as "clipping". The
+    goal is to accept reasonable real-world recordings — not to repair badly
+    distorted ones. True clipping artifacts inside the waveform can't be
+    recovered by peak scaling and will still sound bad.
+
+    Args:
+        audio: Mono audio array.
+        sample_rate: Sample rate of ``audio`` in Hz.
+        peak_target: Peak amplitude cap in [0, 1]. Applied only if the input
+            peak exceeds this value.
+        trim_top_db: Silence threshold for edge trimming, in dB below peak.
+            40 dB sits below normal speech dynamic range (≈30 dB) so soft
+            trailing syllables are preserved, while still catching obvious
+            leading/trailing silence. Lower values are more aggressive;
+            librosa's own default is 60.
+        edge_padding_ms: Milliseconds of padding to add back at each edge
+            *only if* trimming shortened the waveform, so TTS engines have a
+            brief silence to anchor on without ever making the output longer
+            than the input.
+
+    Returns:
+        Preprocessed audio array (float32).
+    """
+    audio = audio.astype(np.float32, copy=False)
+
+    if audio.size == 0:
+        return audio
+
+    audio = audio - float(np.mean(audio))
+
+    trimmed, _ = librosa.effects.trim(audio, top_db=trim_top_db)
+    if 0 < trimmed.size < audio.size:
+        pad_each = int(sample_rate * edge_padding_ms / 1000)
+        # Never pad past the original length — for near-max-duration uploads
+        # an unconditional pad would push them over the 30 s ceiling and
+        # trigger a spurious "too long" rejection.
+        headroom = (audio.size - trimmed.size) // 2
+        pad = min(pad_each, max(headroom, 0))
+        if pad > 0:
+            trimmed = np.pad(trimmed, (pad, pad), mode="constant")
+        audio = trimmed
+
+    peak = float(np.abs(audio).max())
+    if peak > peak_target and peak > 0:
+        audio = audio * (peak_target / peak)
+
+    return audio
+
+
 def validate_reference_audio(
     audio_path: str,
     min_duration: float = 2.0,
@@ -207,13 +267,13 @@ def validate_reference_audio(
 ) -> Tuple[bool, Optional[str]]:
     """
     Validate reference audio for voice cloning.
-    
+
     Args:
         audio_path: Path to audio file
         min_duration: Minimum duration in seconds
         max_duration: Maximum duration in seconds
         min_rms: Minimum RMS level
-        
+
     Returns:
         Tuple of (is_valid, error_message)
     """
@@ -231,26 +291,28 @@ def validate_and_load_reference_audio(
 ) -> Tuple[bool, Optional[str], Optional[np.ndarray], Optional[int]]:
     """
     Validate and load reference audio in a single pass.
-    
+
+    Applies :func:`preprocess_reference_audio` before checks so that
+    slightly-hot recordings aren't rejected as clipping. Duration and RMS
+    checks run on the preprocessed waveform.
+
     Returns:
         Tuple of (is_valid, error_message, audio_array, sample_rate)
     """
     try:
         audio, sr = load_audio(audio_path)
+        audio = preprocess_reference_audio(audio, sr)
         duration = len(audio) / sr
-        
+
         if duration < min_duration:
             return False, f"Audio too short (minimum {min_duration} seconds)", None, None
         if duration > max_duration:
             return False, f"Audio too long (maximum {max_duration} seconds)", None, None
-        
+
         rms = np.sqrt(np.mean(audio**2))
         if rms < min_rms:
             return False, "Audio is too quiet or silent", None, None
-        
-        if np.abs(audio).max() > 0.99:
-            return False, "Audio is clipping (reduce input gain)", None, None
-        
+
         return True, None, audio, sr
     except Exception as e:
         return False, f"Error validating audio: {str(e)}", None, None