NVIDIA#1369 Updated DeepLearningExamples so that all projects can use latest librosa.

xvdp · xvdp · commit 400f37f66751 · 2024-01-16T08:06:01.000-08:00
modified:   CUDA-Optimized/FastSpeech/fastspeech/dataset/ljspeech_dataset.py
	modified:   CUDA-Optimized/FastSpeech/generate.py
	modified:   CUDA-Optimized/FastSpeech/tacotron2/audio_processing.py
	modified:   CUDA-Optimized/FastSpeech/tacotron2/layers.py
	modified:   Kaldi/SpeechRecognition/notebooks/Kaldi_TRTIS_inference_offline_demo.ipynb
	modified:   Kaldi/SpeechRecognition/notebooks/Kaldi_TRTIS_inference_online_demo.ipynb
	modified:   PyTorch/SpeechRecognition/Jasper/requirements.txt
	modified:   PyTorch/SpeechRecognition/QuartzNet/requirements.txt
	modified:   PyTorch/SpeechRecognition/wav2vec2/requirements.txt
	modified:   PyTorch/SpeechSynthesis/FastPitch/hifigan/data_function.py
	modified:   PyTorch/SpeechSynthesis/FastPitch/requirements.txt
	modified:   PyTorch/SpeechSynthesis/HiFiGAN/requirements.txt
	modified:   PyTorch/SpeechSynthesis/Tacotron2/notebooks/conversationalai/client/speech_ai_demo/utils/jasper/speech_utils.py
	modified:   PyTorch/SpeechSynthesis/Tacotron2/trtis_cpp/src/trt/requirements.txt
diff --git a/CUDA-Optimized/FastSpeech/fastspeech/dataset/ljspeech_dataset.py b/CUDA-Optimized/FastSpeech/fastspeech/dataset/ljspeech_dataset.py
@@ -107,7 +107,7 @@ def __getitem__(self, idx):
 
             # Audio processing
             wav, _ = librosa.effects.trim(wav, frame_length=self.win_len, hop_length=self.hop_len)
-            
+
             if self.mels_path:
                 mel = np.load(os.path.join(self.mels_path, name + ".mel.npy"))
             else:
diff --git a/CUDA-Optimized/FastSpeech/generate.py b/CUDA-Optimized/FastSpeech/generate.py
@@ -28,7 +28,7 @@
 import time
 
 import fire
-import librosa
+import soundfile
 import torch
 
 from fastspeech.data_load import PadDataLoader
@@ -158,7 +158,7 @@ def generate(hparam='infer.yaml',
                     wav = wav[:wav_len]
 
                     path = os.path.join(results_path, text[:MAX_FILESIZE] + ".wav")
-                    librosa.output.write_wav(path, wav, hp.sr)
+                    soundfile.write(path, wav, hp.sr)
 
         except StopIteration:
             tprint("Generation has been done.")
diff --git a/CUDA-Optimized/FastSpeech/tacotron2/audio_processing.py b/CUDA-Optimized/FastSpeech/tacotron2/audio_processing.py
@@ -79,7 +79,7 @@ def window_sumsquare(window, n_frames, hop_length=200, win_length=800,
     # Compute the squared window at the desired length
     win_sq = get_window(window, win_length, fftbins=True)
     win_sq = librosa_util.normalize(win_sq, norm=norm)**2
-    win_sq = librosa_util.pad_center(win_sq, n_fft)
+    win_sq = librosa_util.pad_center(win_sq, size=n_fft)
 
     # Fill the envelope
     for i in range(n_frames):
diff --git a/CUDA-Optimized/FastSpeech/tacotron2/layers.py b/CUDA-Optimized/FastSpeech/tacotron2/layers.py
@@ -31,7 +31,6 @@
 """https://github.com/NVIDIA/tacotron2"""
 
 import torch
-from librosa.filters import mel as librosa_mel_fn
 
 
 class LinearNorm(torch.nn.Module):
diff --git a/Kaldi/SpeechRecognition/notebooks/Kaldi_TRTIS_inference_offline_demo.ipynb b/Kaldi/SpeechRecognition/notebooks/Kaldi_TRTIS_inference_offline_demo.ipynb
@@ -511,10 +511,10 @@
     "        \"\"\"\n",
     "        samples = self._convert_samples_to_float32(samples)\n",
     "        if target_sr is not None and target_sr != sample_rate:\n",
-    "            samples = librosa.core.resample(samples, sample_rate, target_sr)\n",
+    "            samples = librosa.core.resample(samples, orig_sr=sample_rate, target_sr=target_sr)\n",
     "            sample_rate = target_sr\n",
     "        if trim:\n",
-    "            samples, _ = librosa.effects.trim(samples, trim_db)\n",
+    "            samples, _ = librosa.effects.trim(samples, top_db=trim_db)\n",
     "        self._samples = samples\n",
     "        self._sample_rate = sample_rate\n",
     "        if self._samples.ndim >= 2:\n",
diff --git a/Kaldi/SpeechRecognition/notebooks/Kaldi_TRTIS_inference_online_demo.ipynb b/Kaldi/SpeechRecognition/notebooks/Kaldi_TRTIS_inference_online_demo.ipynb
@@ -640,10 +640,10 @@
     "        \"\"\"\n",
     "        samples = self._convert_samples_to_float32(samples)\n",
     "        if target_sr is not None and target_sr != sample_rate:\n",
-    "            samples = librosa.core.resample(samples, sample_rate, target_sr)\n",
+    "            samples = librosa.core.resample(samples, orig_sr=sample_rate, target_sr=target_sr)\n",
     "            sample_rate = target_sr\n",
     "        if trim:\n",
-    "            samples, _ = librosa.effects.trim(samples, trim_db)\n",
+    "            samples, _ = librosa.effects.trim(samples, top_db=trim_db)\n",
     "        self._samples = samples\n",
     "        self._sample_rate = sample_rate\n",
     "        if self._samples.ndim >= 2:\n",
diff --git a/PyTorch/SpeechRecognition/Jasper/requirements.txt b/PyTorch/SpeechRecognition/Jasper/requirements.txt
@@ -1,6 +1,6 @@
 inflect==5.3.0
 ipdb
-librosa==0.9.0
+librosa>=0.9.0
 pandas==1.5.2
 pyyaml>=5.4
 soundfile
diff --git a/PyTorch/SpeechRecognition/QuartzNet/requirements.txt b/PyTorch/SpeechRecognition/QuartzNet/requirements.txt
@@ -1,6 +1,6 @@
 inflect==5.3.0
 ipdb
-librosa==0.9.0
+librosa>=0.9.0
 pandas==1.5.2
 pyyaml>=5.4
 soundfile
diff --git a/PyTorch/SpeechRecognition/wav2vec2/requirements.txt b/PyTorch/SpeechRecognition/wav2vec2/requirements.txt
@@ -1,5 +1,5 @@
 editdistance==0.6.0
-librosa==0.10.1
+librosa>=0.10.1
 omegaconf==2.0.6  # optional for handling certain Fairseq ckpts
 pyarrow==6.0.1
 soundfile==0.12.1
diff --git a/PyTorch/SpeechSynthesis/FastPitch/hifigan/data_function.py b/PyTorch/SpeechSynthesis/FastPitch/hifigan/data_function.py
@@ -69,7 +69,7 @@ def mel_spectrogram(y, n_fft, num_mels, sampling_rate, hop_size, win_size,
     global mel_basis, hann_window
     fmax_key = f'{fmax}_{y.device}'
     if fmax_key not in mel_basis:
-        mel = librosa_mel_fn(sampling_rate, n_fft, num_mels, fmin, fmax)
+        mel = librosa_mel_fn(sr=sampling_rate, n_fft=n_fft, n_mels=num_mels, fmin=fmin, fmax=fmax)
         mel_basis[fmax_key] = torch.from_numpy(mel).float().to(y.device)
         hann_window[str(y.device)] = torch.hann_window(win_size).to(y.device)
 
diff --git a/PyTorch/SpeechSynthesis/FastPitch/requirements.txt b/PyTorch/SpeechSynthesis/FastPitch/requirements.txt
@@ -1,5 +1,5 @@
 inflect
-librosa==0.9.0
+librosa>=0.9.0
 matplotlib
 numpy
 pynvml==11.0.0
diff --git a/PyTorch/SpeechSynthesis/HiFiGAN/requirements.txt b/PyTorch/SpeechSynthesis/HiFiGAN/requirements.txt
@@ -1,5 +1,5 @@
 inflect
-librosa==0.9.0
+librosa>=0.9.0
 numpy
 pandas
 pynvml==11.0.0
diff --git a/PyTorch/SpeechSynthesis/Tacotron2/notebooks/conversationalai/client/speech_ai_demo/utils/jasper/speech_utils.py b/PyTorch/SpeechSynthesis/Tacotron2/notebooks/conversationalai/client/speech_ai_demo/utils/jasper/speech_utils.py
@@ -383,10 +383,10 @@ def __init__(self, samples, sample_rate, target_sr=16000, trim=False,
         """
         samples = self._convert_samples_to_float32(samples)
         if target_sr is not None and target_sr != sample_rate:
-            samples = librosa.core.resample(samples, sample_rate, target_sr)
+            samples = librosa.core.resample(samples, orig_sr=sample_rate, target_sr=target_sr)
             sample_rate = target_sr
         if trim:
-            samples, _ = librosa.effects.trim(samples, trim_db)
+            samples, _ = librosa.effects.trim(samples, top_db=trim_db)
         self._samples = samples
         self._sample_rate = sample_rate
         if self._samples.ndim >= 2:
diff --git a/PyTorch/SpeechSynthesis/Tacotron2/trtis_cpp/src/trt/requirements.txt b/PyTorch/SpeechSynthesis/Tacotron2/trtis_cpp/src/trt/requirements.txt
@@ -1,4 +1,4 @@
 torch==1.3.0
 onnx==1.5.0
 scipy==1.3.1
-librosa==0.7.0
+librosa
