init

fetch_dataset_s2s.py

@@ -135,13 +135,10 @@ def get_audio(dataframe: pd.DataFrame):
                     wav = wav[:int(end/sampling_rate * sr) + sr]
                     if len(wav) < min_seq_length or len(wav) > max_seq_length:
                         print(f"invalid wave length: {len(wav)}")
+                        cleanup(features, feature_file)
                         return None
                     print(f"SAVING: {features[f'{side}.path']}")
                     sf.write(features[f"{side}.path"], wav, sr)
-                    # if sr != sampling_rate:
-                    #     print(f"RESAMPLING: {wav.shape} length audio")
-                    #     wav = librosa.resample(wav, orig_sr=sr, target_sr=sampling_rate)
-                    # sf.write(features[f"{side}.path"], wav[start:end], sampling_rate)
 
                 except Exception as e:
                     print(f"\n#### ERROR ####\n {e}")

main_s2s.sh

@@ -312,8 +312,8 @@ done
 # speaker embedding
 export SE_MODEL="metavoice"
 export SE_MODEL="w2vbert-600m"
-export SE_MODEL="xlsr-2b"
 export SE_MODEL="hubert-xl"
+export SE_MODEL="xlsr-2b"
 for i in $(seq 1 100);
 do
   export DATASET_ID=${i}