Open-source speech recognition toolkit for training, inference, streaming ASR, VAD, punctuation, speaker diarization pipelines, and OpenAI-compatible/MCP serving.
Python asraudiochineseemotion-recognitionfunasrmcp-servermultilingual-asropenai-compatible-apiparaformerpunctuationpytorchreal-time-asrspeaker-diarizationspeech-recognitionspeech-to-textstreaming-asrtranscriptionvllmvoice-activity-detectionwhisper-alternative
fix: align randperm index device in paraformer streaming sampler
May131176e623f9
diff --git a/funasr/models/paraformer_streaming/model.py b/funasr/models/paraformer_streaming/model.pyindex 4f56c9a9b..287ffdfae 100644--- a/funasr/models/paraformer_streaming/model.py+++ b/funasr/models/paraformer_streaming/model.py@@ -377,8 +377,11 @@ class ParaformerStreaming(Paraformer):((seq_lens[li] - same_num[li].sum()).float()) * self.sampling_ratio).long()if target_num > 0:+ index = torch.randperm(+ seq_lens[li].item(), device=input_mask.device+ )[: target_num.item()]input_mask[li].scatter_(- dim=0, index=torch.randperm(seq_lens[li])[:target_num].cuda(), value=0+ dim=0, index=index, value=0)input_mask = input_mask.eq(1)input_mask = input_mask.masked_fill(~nonpad_positions, False)