Fix: use soundfile for audio I/O (avoid torchcodec), add attention_mask for TTS, add soundfile dep

app.py

@@ -6,6 +6,7 @@ from pathlib import Path
 
 import gradio as gr
 import torch
+import soundfile as sf
 import torchaudio
 from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
 from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
@@ -46,17 +47,21 @@ def load_asr():
 
 def transcribe(wav):
     processor, model = load_asr()
-    audio, sr = torchaudio.load(str(wav))
-    audio = audio.mean(0)
+    # Use soundfile to load (avoids torchcodec issues)
+    audio, sr = sf.read(str(wav))
+    if len(audio.shape) > 1:
+        audio = audio.mean(axis=1)
     if sr != 16000:
-        audio = torchaudio.functional.resample(audio, sr, 16000)
+        audio = torchaudio.functional.resample(
+            torch.from_numpy(audio).float(), sr, 16000
+        ).numpy()
 
-    total = audio.numel() / 16000
+    total = len(audio) / 16000
     pieces = []
     chunk = 30
 
     for start in range(0, int(total + 0.999), chunk):
-        x = audio[start*16000:min(int(total*16000),(start+chunk)*16000)].numpy()
+        x = audio[start*16000:min(int(total*16000),(start+chunk)*16000)]
         if len(x) < 400:
             continue
         inputs = processor(x, sampling_rate=16000, return_tensors="pt")
@@ -107,10 +112,12 @@ def load_tts():
 
 def tts(text, wav):
     tokenizer, model = load_tts()
-    inputs = tokenizer(text, return_tensors="pt")
+    # Add attention_mask to avoid warning
+    inputs = tokenizer(text, return_tensors="pt", padding=True)
     with torch.inference_mode():
         audio = model(**inputs).waveform.cpu()
-    torchaudio.save(str(wav), audio, model.config.sampling_rate)
+    # Use soundfile instead of torchaudio.save (avoids torchcodec)
+    sf.write(str(wav), audio.squeeze().numpy(), model.config.sampling_rate)
 
 def fit_audio(src, dst, target):
     target = max(float(target), 0.1)

requirements.txt

@@ -5,3 +5,4 @@ transformers>=4.48,<5
 sentencepiece>=0.2
 safetensors>=0.4
 numpy<2.3
+soundfile>=0.12