Add some VAD to the pytorch speech-to-text example.

scripts/stt_from_file_pytorch.py

@@ -4,7 +4,7 @@
 #     "julius",
 #     "librosa",
 #     "soundfile",
-#     "moshi",
+#     "moshi==0.2.9",
 # ]
 # ///
 
@@ -20,8 +20,8 @@ import math
 import julius
 import moshi.models
 import sphn
+import time
 import torch
-import tqdm
 
 
 @dataclasses.dataclass
@@ -138,9 +138,13 @@ def main(args):
 
     mimi = info.get_mimi(device=args.device)
     tokenizer = info.get_text_tokenizer()
+    lm_kwargs_overrides = {}
+    if args.vad:
+        lm_kwargs_overrides = {"extra_heads_num_heads": 4}
     lm = info.get_moshi(
         device=args.device,
         dtype=torch.bfloat16,
+        lm_kwargs_overrides=lm_kwargs_overrides,
     )
     lm_gen = moshi.models.LMGen(lm, temp=0, temp_text=0.0)
 
@@ -171,14 +175,35 @@ def main(args):
         itertools.repeat(silence_chunk, n_suffix_chunks),
     )
 
+    start_time = time.time()
+    nchunks = 0
+    last_print_was_vad = False
     with mimi.streaming(1), lm_gen.streaming(1):
-        for audio_chunk in tqdm.tqdm(chunks):
+        for audio_chunk in chunks:
+            nchunks += 1
             audio_tokens = mimi.encode(audio_chunk)
+            if args.vad:
+                text_tokens, vad_heads = lm_gen.step_with_extra_heads(audio_tokens)
+                if vad_heads:
+                    pr_vad = vad_heads[2][0, 0, 0].cpu().item()
+                    if pr_vad > 0.5 and not last_print_was_vad:
+                        print(" [end of turn detected]")
+                        last_print_was_vad = True
+            else:
                 text_tokens = lm_gen.step(audio_tokens)
-            if text_tokens is not None:
+            text_token = text_tokens[0, 0, 0].cpu().item()
+            if text_token not in (0, 3):
+                _text = tokenizer.id_to_piece(text_tokens[0, 0, 0].cpu().item())  # type: ignore
+                _text = _text.replace("▁", " ")
+                print(_text, end="", flush=True)
+                last_print_was_vad = False
             text_tokens_accum.append(text_tokens)
 
     utterance_tokens = torch.concat(text_tokens_accum, dim=-1)
+    dt = time.time() - start_time
+    print(
+        f"\nprocessed {nchunks} chunks in {dt:.2f} seconds, steps per second: {nchunks / dt:.2f}"
+    )
     timed_text = tokens_to_timestamped_text(
         utterance_tokens,
         tokenizer,
@@ -209,6 +234,9 @@ if __name__ == "__main__":
     parser.add_argument(
         "--config-path", type=str, help="Path to a local config file.", default=None
     )
+    parser.add_argument(
+        "--vad", action="store_true", help="Enable VAD (Voice Activity Detection)."
+    )
     parser.add_argument(
         "--device",
         type=str,