[demo] refine app.py

app.py

@@ -14,10 +14,7 @@
 # limitations under the License.
 
 import gradio as gr
-import torchaudio
-import torchaudio.compliance.kaldi as kaldi
-import torch
-import onnxruntime as ort
+import wespeakerruntime as wespeaker
 from sklearn.metrics.pairwise import cosine_similarity
 
 STYLE = """
@@ -49,53 +46,9 @@ OUTPUT_ERROR = (STYLE + """
     </div>
 """)
 
+vox_model = wespeaker.Inference('pre_model/voxceleb_resnet34.onnx')
+cnc_model = wespeaker.Inference('pre_model/cnceleb_resnet34.onnx')
 
-def compute_fbank(wav_path,
-                  num_bel_bins=80,
-                  frame_length=25,
-                  frame_shift=10,
-                  dither=0.0,
-                  resample_rate=16000):
-    """ Extract fbank, simlilar to the one in wespeaker.dataset.processor,
-        While integrating the wave reading and CMN.
-    """
-    waveform, sample_rate = torchaudio.load(wav_path)
-    # resample
-    if sample_rate != resample_rate:
-        waveform = torchaudio.transforms.Resample(
-            orig_freq=sample_rate, new_freq=resample_rate)(waveform)
-    waveform = waveform * (1 << 15)
-    mat = kaldi.fbank(waveform,
-                      num_mel_bins=num_bel_bins,
-                      frame_length=frame_length,
-                      frame_shift=frame_shift,
-                      dither=dither,
-                      sample_frequency=sample_rate,
-                      window_type='hamming',
-                      use_energy=False)
-    # CMN, without CVN
-    mat = mat - torch.mean(mat, dim=0)
-    return mat
-
-
-class OnnxModel(object):
-
-    def __init__(self, model_path):
-        so = ort.SessionOptions()
-        so.inter_op_num_threads = 1
-        so.intra_op_num_threads = 1
-        self.session = ort.InferenceSession(model_path, sess_options=so)
-
-    def extract_embedding(self, wav_path):
-        feats = compute_fbank(wav_path)
-        feats = feats.unsqueeze(0).numpy()
-
-        embeddings = self.session.run(output_names=['embs'],
-                                      input_feed={'feats': feats})
-        return embeddings[0]
-
-vox_model = OnnxModel('pre_model/voxceleb_resnet34.onnx')
-cnc_model = OnnxModel('pre_model/cnceleb_resnet34.onnx')
 
 def speaker_verification(audio_path1, audio_path2, lang='CN'):
     if audio_path1 == None or audio_path2 == None:
@@ -108,13 +61,13 @@ def speaker_verification(audio_path1, audio_path2, lang='CN'):
     else:
         output = OUTPUT_ERROR.format('Please select a language')
         return output
-    emb1 = model.extract_embedding(audio_path1)
-    emb2 = model.extract_embedding(audio_path2)
+    emb1 = model.extract_embedding_wav(audio_path1)
+    emb2 = model.extract_embedding_wav(audio_path2)
     cos_score = cosine_similarity(emb1.reshape(1, -1), emb2.reshape(1,
                                                                     -1))[0][0]
     cos_score = (cos_score + 1) / 2.0
 
-    if cos_score >= 0.73:
+    if cos_score >= 0.70:
         output = OUTPUT_OK.format(cos_score * 100)
     else:
         output = OUTPUT_FAIL.format(cos_score * 100)

requirements.txt

@@ -1,4 +1,3 @@
-onnxruntime==1.11.1
 gradio
-torchaudio
-scikit-learn
+wespeakerruntime
+scikit-learn