update asr_spk inference for shot utt

shixian.shi

2023-11-24 72fecc8e038070affbf223f8965871e8a1c8c001

 funasr/bin/asr_inference_launch.py

@@ -20,7 +20,8 @@
import numpy as np
import torch
import torchaudio
import soundfile
# import librosa
import librosa
import yaml

from funasr.bin.asr_infer import Speech2Text
@@ -955,24 +956,29 @@
                ed = int(vadsegment[1]) / 1000
                vad_segments.append(
                    [st, ed, audio[int(st * 16000):int(ed * 16000)]])
            check_audio_list(vad_segments)
            # sv pipeline
            segments = sv_chunk(vad_segments)
            embeddings = []
            for s in segments:
                #_, embs = self.sv_pipeline([s[2]], output_emb=True)
                # embeddings.append(embs)
                wavs = sv_preprocess([s[2]])
                # embs = self.forward(wavs)
                embs = []
                for x in wavs:
                    x = extract_feature([x])
                    embs.append(sv_model(x))
                embs = torch.cat(embs)
                embeddings.append(embs.detach().numpy())
            embeddings = np.concatenate(embeddings)
            labels = cb_model(embeddings)
            sv_output = postprocess(segments, vad_segments, labels, embeddings)
            audio_dur = check_audio_list(vad_segments)
            if audio_dur > 5:
                # sv pipeline
                segments = sv_chunk(vad_segments)
                embeddings = []
                for s in segments:
                    #_, embs = self.sv_pipeline([s[2]], output_emb=True)
                    # embeddings.append(embs)
                    wavs = sv_preprocess([s[2]])
                    # embs = self.forward(wavs)
                    embs = []
                    for x in wavs:
                        x = extract_feature([x])
                        embs.append(sv_model(x))
                    embs = torch.cat(embs)
                    embeddings.append(embs.detach().numpy())
                embeddings = np.concatenate(embeddings)
                labels = cb_model(embeddings)
                sv_output = postprocess(segments, vad_segments, labels, embeddings)
            else:
                # fake speaker res for too shot utterance
                sv_output = [[0.0, vadsegments[-1][-1]/1000.0, 0]]
                logging.warning("Too short utterence found: {}, return default speaker results.".format(keys))

            speech, speech_lengths = batch["speech"], batch["speech_lengths"]

@@ -1281,7 +1287,8 @@
            try:
                raw_inputs = torchaudio.load(data_path_and_name_and_type[0])[0][0]
            except:
                raw_inputs = soundfile.read(data_path_and_name_and_type[0], dtype='float32')[0]
                # raw_inputs = librosa.load(data_path_and_name_and_type[0], dtype='float32')[0]
                raw_inputs, sr = librosa.load(data_path_and_name_and_type[0], dtype='float32')
                if raw_inputs.ndim == 2:
                    raw_inputs = raw_inputs[:, 0]
                raw_inputs = torch.tensor(raw_inputs)