成功推理 完全正确（用了高级 Debug 技巧，只有一条训练数据）

examples/asr/transcribe_speech.py

@@ -32,7 +32,6 @@ from nemo.collections.asr.parts.utils.eval_utils import cal_write_wer
 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis
 from nemo.collections.asr.parts.utils.transcribe_utils import (
     compute_output_filename,
-    get_inference_dtype,
     prepare_audio_data,
     restore_transcription_order,
     setup_model,
@@ -276,12 +275,10 @@ def main(cfg: TranscriptionConfig) -> Union[TranscriptionConfig, List[Hypothesis
 
     amp_dtype = torch.float16 if cfg.amp_dtype == "float16" else torch.bfloat16
 
-    compute_dtype: torch.dtype
     if cfg.amp:
-        # with amp model weights required to be in float32
-        compute_dtype = torch.float32
+        compute_dtype = torch.float16 if cfg.amp_dtype == "float16" else torch.bfloat16
     else:
-        compute_dtype = get_inference_dtype(compute_dtype=cfg.compute_dtype, device=map_location)
+        compute_dtype = torch.float32
 
     asr_model.to(compute_dtype)
 
@@ -358,15 +355,20 @@ def main(cfg: TranscriptionConfig) -> Union[TranscriptionConfig, List[Hypothesis
         else:
             cfg.decoding = cfg.rnnt_decoding
 
+    # Compute output filename
+    cfg = compute_output_filename(cfg, model_name)
+
     filepaths, sorted_manifest_path = prepare_audio_data(cfg)
 
+    # 如果 audio_dir 本身就是一个文件，则直接使用该文件
+    if cfg.audio_dir is not None and os.path.isfile(cfg.audio_dir):
+        filepaths = [cfg.audio_dir]
+        sorted_manifest_path = None
+
     remove_path_after_done = sorted_manifest_path if sorted_manifest_path is not None else None
 
     filepaths = sorted_manifest_path if sorted_manifest_path is not None else filepaths
 
-    # Compute output filename
-    cfg = compute_output_filename(cfg, model_name)
-
     # if transcripts should not be overwritten, and already exists, skip re-transcription step and return
     if not cfg.return_transcriptions and not cfg.overwrite_transcripts and os.path.exists(cfg.output_filename):
         logging.info(
@@ -466,6 +468,19 @@ def main(cfg: TranscriptionConfig) -> Union[TranscriptionConfig, List[Hypothesis
         compute_langs=compute_langs,
         timestamps=cfg.timestamps,
     )
+    
+    # 强制将输出的 JSON 文件中的 unicode 转为 utf-8 字符
+    if os.path.exists(output_filename):
+        with open(output_filename, 'r', encoding='utf-8') as f:
+            lines = f.readlines()
+        with open(output_filename, 'w', encoding='utf-8') as f:
+            for line in lines:
+                try:
+                    data = json.loads(line)
+                    f.write(json.dumps(data, ensure_ascii=False) + '\n')
+                except:
+                    f.write(line)
+
     logging.info(f"Finished writing predictions to {output_filename}!")
 
     # clean-up
@@ -498,4 +513,16 @@ def main(cfg: TranscriptionConfig) -> Union[TranscriptionConfig, List[Hypothesis
 
 
 if __name__ == '__main__':
+
+    import sys
+    sys.argv.extend([
+        "model_path=results/NeMo_Ja/checkpoints/NeMo_Ja.nemo",
+        "audio_dir=E:/huggingface_echodict/NeMo/data/common_voice_11_0/ja/train/ja_train_0/common_voice_ja_25372057.wav",
+        "output_filename=transcripts.json",
+        "batch_size=16",
+        "amp=True",
+        "cuda=0",
+
+    ])
+    
     main()  # noqa pylint: disable=no-value-for-parameter

readme.txt

@@ -220,9 +220,11 @@ python examples/asr/asr_ctc/speech_to_text_ctc_bpe.py \
   model.test_ds.manifest_filepath=$TEST_MANIFEST
 
 
-  python examples\asr\asr_ctc\speech_to_text_ctc_bpe.py
-    开始训练
+开始训练
+python examples\asr\asr_ctc\speech_to_text_ctc_bpe.py
 
+成功推理 完全正确（用了高级 Debug 技巧，只有一条训练数据）
+python examples/asr/transcribe_speech.py