[TTS] Fix TTS audio preprocessing bugs (NVIDIA#6628)

Signed-off-by: Ryan <[email protected]>

Author: rlangman

nemo/collections/tts/parts/preprocessing/features.py

@@ -131,10 +131,14 @@ def __init__(
             n_fft=win_length,
             lowfreq=lowfreq,
             highfreq=highfreq,
+            mag_power=1.0,
             log=log,
             log_zero_guard_type=log_zero_guard_type,
             log_zero_guard_value=log_zero_guard_value,
             mel_norm=mel_norm,
+            normalize=None,
+            preemph=None,
+            dither=0.0,
         )
 
     def compute_mel_spec(self, manifest_entry: dict, audio_dir: Path) -> Tensor:

nemo/collections/tts/parts/utils/tts_dataset_utils.py

@@ -67,6 +67,9 @@ def normalize_volume(audio: np.array, volume_level: float) -> np.array:
     if not (0.0 <= volume_level <= 1.0):
         raise ValueError(f"Volume must be in range [0.0, 1.0], received {volume_level}")
 
+    if audio.size == 0:
+        return audio
+
     max_sample = np.max(np.abs(audio))
     if max_sample == 0:
         return audio

scripts/dataset_processing/tts/audio_processing/preprocess_audio.py

@@ -128,7 +128,7 @@ def _process_entry(
     if audio_trimmer is not None:
         audio, start_i, end_i = audio_trimmer.trim_audio(audio=audio, sample_rate=sample_rate, audio_id=audio_path)
 
-    if output_sample_rate is not None:
+    if output_sample_rate:
         audio = librosa.resample(y=audio, orig_sr=sample_rate, target_sr=output_sample_rate)
         sample_rate = output_sample_rate
 
@@ -140,7 +140,7 @@ def _process_entry(
     original_duration = librosa.get_duration(filename=audio_path)
     output_duration = librosa.get_duration(filename=output_path)
 
-    entry["duration"] = output_duration
+    entry["duration"] = round(output_duration, 2)
 
     if os.path.isabs(audio_filepath):
         entry["audio_filepath"] = output_path

tests/collections/tts/data/test_data_utils.py

@@ -14,9 +14,10 @@
 
 from pathlib import Path
 
+import numpy as np
 import pytest
 
-from nemo.collections.tts.parts.utils.tts_dataset_utils import get_abs_rel_paths, get_audio_filepaths
+from nemo.collections.tts.parts.utils.tts_dataset_utils import get_abs_rel_paths, get_audio_filepaths, normalize_volume
 
 
 class TestTTSDatasetUtils:
@@ -53,3 +54,68 @@ def test_get_audio_paths(self):
 
         assert abs_path == Path("/home/audio/examples/example.wav")
         assert rel_path == audio_rel_path
+
+    @pytest.mark.run_only_on('CPU')
+    @pytest.mark.unit
+    def test_normalize_volume(self):
+        input_audio = np.array([0.0, 0.1, 0.3, 0.5])
+        expected_output = np.array([0.0, 0.18, 0.54, 0.9])
+
+        output_audio = normalize_volume(audio=input_audio, volume_level=0.9)
+
+        np.testing.assert_array_almost_equal(output_audio, expected_output)
+
+    @pytest.mark.run_only_on('CPU')
+    @pytest.mark.unit
+    def test_normalize_volume_negative_peak(self):
+        input_audio = np.array([0.0, 0.1, -0.3, -1.0, 0.5])
+        expected_output = np.array([0.0, 0.05, -0.15, -0.5, 0.25])
+
+        output_audio = normalize_volume(audio=input_audio, volume_level=0.5)
+
+        np.testing.assert_array_almost_equal(output_audio, expected_output)
+
+    @pytest.mark.run_only_on('CPU')
+    @pytest.mark.unit
+    def test_normalize_volume_zero(self):
+        input_audio = np.array([0.0, 0.1, 0.3, 0.5])
+        expected_output = np.array([0.0, 0.0, 0.0, 0.0])
+
+        output_audio = normalize_volume(audio=input_audio, volume_level=0.0)
+
+        np.testing.assert_array_almost_equal(output_audio, expected_output)
+
+    @pytest.mark.run_only_on('CPU')
+    @pytest.mark.unit
+    def test_normalize_volume_max(self):
+        input_audio = np.array([0.0, 0.1, 0.3, 0.5])
+        expected_output = np.array([0.0, 0.2, 0.6, 1.0])
+
+        output_audio = normalize_volume(audio=input_audio, volume_level=1.0)
+
+        np.testing.assert_array_almost_equal(output_audio, expected_output)
+
+    @pytest.mark.run_only_on('CPU')
+    @pytest.mark.unit
+    def test_normalize_volume_zeros(self):
+        input_audio = np.array([0.0, 0.0, 0.0])
+
+        output_audio = normalize_volume(audio=input_audio, volume_level=0.5)
+
+        np.testing.assert_array_almost_equal(output_audio, input_audio)
+
+    @pytest.mark.run_only_on('CPU')
+    @pytest.mark.unit
+    def test_normalize_volume_empty(self):
+        input_audio = np.array([])
+
+        output_audio = normalize_volume(audio=input_audio, volume_level=1.0)
+
+        np.testing.assert_array_almost_equal(output_audio, input_audio)
+
+    @pytest.mark.run_only_on('CPU')
+    @pytest.mark.unit
+    def test_normalize_volume_out_of_range(self):
+        input_audio = np.array([0.0, 0.1, 0.3, 0.5])
+        with pytest.raises(ValueError, match="Volume must be in range"):
+            normalize_volume(audio=input_audio, volume_level=2.0)