CTC 训练推理正常，for nemo==2.8.0.rc0，但是推理时第一个字符被吃了，其他都正常

examples/asr/asr_cache_aware_streaming/speech_to_text_cache_aware_streaming_infer.py

@@ -115,7 +115,8 @@ from nemo.collections.asr.parts.submodules.rnnt_decoding import RNNTDecodingConf
 from nemo.collections.asr.parts.utils.manifest_utils import read_manifest
 from nemo.collections.asr.parts.utils.rnnt_utils import Hypothesis
 from nemo.collections.asr.parts.utils.streaming_utils import CacheAwareStreamingAudioBuffer
-from nemo.collections.asr.parts.utils.transcribe_utils import setup_model, normalize_timestamp_output, process_timestamp_outputs
+from nemo.collections.asr.parts.utils.transcribe_utils import setup_model, normalize_timestamp_output
+from nemo.collections.asr.parts.utils.timestamp_utils import process_timestamp_outputs
 from nemo.core.config import hydra_runner
 from nemo.utils import logging
 

examples/asr/asr_ctc/speech_to_text_ctc_bpe.py

@@ -94,7 +94,10 @@ def main(cfg):
 
     if hasattr(cfg.model, 'test_ds') and cfg.model.test_ds.manifest_filepath is not None:
         if asr_model.prepare_test(trainer):
-            trainer.test(asr_model)
+            if trainer.checkpoint_callback and trainer.checkpoint_callback.best_model_path:
+                trainer.test(asr_model, ckpt_path='best')
+            else:
+                trainer.test(asr_model)
 
 
 if __name__ == '__main__':
@@ -109,7 +112,7 @@ if __name__ == '__main__':
         'exp_manager.resume_ignore_no_checkpoint=true',
         'exp_manager.exp_dir=results/',
         'exp_manager.checkpoint_callback_params.save_top_k=1',
-        '++trainer.check_val_every_n_epoch=350',
+        '++trainer.check_val_every_n_epoch=700',
         '++model.encoder.conv_norm_type=layer_norm',
         'model.tokenizer.dir=data/common_voice_11_0/ja/tokenizers/tokenizer_spe_bpe_v2491',
         'model.train_ds.tarred_audio_filepaths=data/common_voice_11_0/ja/train_tarred_1bk/audio__OP_0..1023_CL_.tar',
@@ -119,7 +122,7 @@ if __name__ == '__main__':
         'model.train_ds.manifest_filepath=data/common_voice_11_0/ja/train_tarred_1bk/tarred_audio_manifest.json',
         'model.validation_ds.manifest_filepath=data/common_voice_11_0/ja/validation/validation_common_voice_11_0_manifest.json',
         'model.test_ds.manifest_filepath=data/common_voice_11_0/ja/test/test_common_voice_11_0_manifest.json',
-        'trainer.max_epochs=350',
+        'trainer.max_epochs=700',
         'trainer.devices=1',
         'trainer.accelerator=gpu',
         'trainer.strategy=auto',

nemo/collections/asr/data/audio_to_text.py

@@ -936,7 +936,8 @@ class _TarredAudioToTextDataset(IterableDataset):
                     self.current_bytes, self.current_fn = next(self.iterator)
                     self.offset_id = 0
                 else:
-                    offset_list = self.collection.mapping[self.current_fn]
+                    file_id, _ = os.path.splitext(os.path.basename(self.current_fn))
+                    offset_list = self.collection.mapping[file_id]
                     if len(offset_list) == self.offset_id + 1:
                         self.current_bytes, self.current_fn = next(self.iterator)
                         self.offset_id = 0

readme.txt

@@ -15,6 +15,9 @@ see https://github.com/NVIDIA-NeMo/NeMo/discussions/8473
 
 
 CTC 能正常训练的 nemo 版本是 2.2.1
+common_voice_11_0 需要的版本是 datasets==3.6.0
+numpy==1.26.4
+
 
 
 这个 KeyError 报错是由于 WebDataset 解析 tar 包中的路径与 NeMo 的 Manifest 映射字典键值不匹配 导致的。

requirements.txt

@@ -12,3 +12,8 @@ modal>=1.3.0.post1
 torch==2.11.0
 nemo-toolkit[asr,cu13]
 lhotse
+nv_one_logger_core>=2.3.1
+nv_one_logger_training_telemetry>=2.3.1
+nv_one_logger_pytorch_lightning_integration>=2.3.1
+kaldi-python-io
+kaldialign

server_aware_streaming.py

@@ -13,7 +13,8 @@ import torch
 from aiohttp import web, WSMsgType
 from loguru import logger
 
-from nemo.collections.asr.parts.utils.transcribe_utils import normalize_timestamp_output, process_timestamp_outputs
+from nemo.collections.asr.parts.utils.transcribe_utils import normalize_timestamp_output
+from nemo.collections.asr.parts.utils.timestamp_utils import process_timestamp_outputs
 
 # Enable debug logging with DEBUG_ASR=1
 DEBUG_ASR = os.environ.get("DEBUG_ASR", "0") == "1"

test_websocket_client.py

@@ -166,7 +166,7 @@ if __name__ == "__main__":
     import sys
     # sys.argv.append( '--all' )
 
-    audio_path = "./harvard_16k.wav" # "./common_voice_ja_25372057.wav" # "./harvard_16k.wav"
+    audio_path = "./common_voice_ja_25372057.wav" # "./common_voice_ja_25372057.wav" # "./harvard_16k.wav"
 
     server_url = "ws://127.0.0.1:8080"
 
@@ -174,7 +174,7 @@ if __name__ == "__main__":
     if "--all" in sys.argv:
         asyncio.run(test_multiple_chunk_sizes(audio_path, server_url))
     else:
-        chunk_ms = 500
+        chunk_ms = 700
         if "--chunk" in sys.argv:
             idx = sys.argv.index("--chunk")
             if idx + 1 < len(sys.argv):