From 6fff82b91747ececa88065c8e18d8508f99935c3 Mon Sep 17 00:00:00 2001 From: "LIyong.Guo" <839019390@qq.com> Date: Fri, 27 May 2022 11:23:43 +0800 Subject: [PATCH 1/2] Update egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py Co-authored-by: Fangjun Kuang --- .../ASR/pruned_transducer_stateless6/hubert_xlarge.py | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py b/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py index e25f9b0af..8b609fdc5 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py @@ -178,9 +178,7 @@ class HubertXlargeFineTuned: N = encoder_embedding.shape[0] assert len(cut_list) == N # 320 is from: 16,000 / 50 = sample_rate / hbuert output frame rate - num_frames = [ - supervisions["num_samples"][i].item() // 320 for i in range(N) - ] + num_frames = (supervisions["num_samples"] // 320).tolist() return encoder_embedding, num_frames def ctc_greedy_search(self, batch): From 6e1a92ac7cd0ce16c71a87f8ba75d70c751992b6 Mon Sep 17 00:00:00 2001 From: "LIyong.Guo" <839019390@qq.com> Date: Fri, 27 May 2022 11:24:30 +0800 Subject: [PATCH 2/2] Update egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py Co-authored-by: Fangjun Kuang --- .../ASR/pruned_transducer_stateless6/hubert_xlarge.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py b/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py index 8b609fdc5..4c00c4739 100644 --- a/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py +++ b/egs/librispeech/ASR/pruned_transducer_stateless6/hubert_xlarge.py @@ -68,7 +68,7 @@ class HubertXlargeFineTuned: """ A wrapper of hubert extra large fine-tuned model. - A teacher model responsible for: + A teacher model is responsible for: 1. load teacher model 2. extracting embeddings to train quantizer. 3. extract codebook indices