Update vad_inference_launch.py

lingyunfly

2023-05-18 19f4fae784210e85421ae2f8dcd0fbbd1eb2ad3e

 funasr/bin/asr_infer.py

@@ -762,23 +762,6 @@
                feats_len = speech_lengths

            if feats.shape[1] != 0:
                if cache_en["is_final"]:
                    if feats.shape[1] + cache_en["chunk_size"][2] < cache_en["chunk_size"][1]:
                        cache_en["last_chunk"] = True
                    else:
                        # first chunk
                        feats_chunk1 = feats[:, :cache_en["chunk_size"][1], :]
                        feats_len = torch.tensor([feats_chunk1.shape[1]])
                        results_chunk1 = self.infer(feats_chunk1, feats_len, cache)

                        # last chunk
                        cache_en["last_chunk"] = True
                        feats_chunk2 = feats[:, -(feats.shape[1] + cache_en["chunk_size"][2] - cache_en["chunk_size"][1]):, :]
                        feats_len = torch.tensor([feats_chunk2.shape[1]])
                        results_chunk2 = self.infer(feats_chunk2, feats_len, cache)

                        return [" ".join(results_chunk1 + results_chunk2)]

                results = self.infer(feats, feats_len, cache)

        return results
@@ -1598,7 +1581,7 @@
            d = ModelDownloader()
            kwargs.update(**d.download_and_unpack(model_tag))
        
        return Speech2Text(**kwargs)
        return Speech2TextTransducer(**kwargs)


class Speech2TextSAASR:
@@ -1640,7 +1623,7 @@
        assert check_argument_types()
        
        # 1. Build ASR model
        from funasr.modules.beam_search.beam_search_sa_asr import BeamSearch
        from funasr.tasks.sa_asr import ASRTask
        scorers = {}
        asr_model, asr_train_args = ASRTask.build_model_from_file(
            asr_train_config, asr_model_file, cmvn_file, device
@@ -1682,6 +1665,7 @@
        # 4. Build BeamSearch object
        # transducer is not supported now
        beam_search_transducer = None
        from funasr.modules.beam_search.beam_search_sa_asr import BeamSearch
        
        weights = dict(
            decoder=1.0 - ctc_weight,