From 219c2482ab755fbd4e49dfbdee91bf1a8a4ec49a Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 19 五月 2023 11:33:27 +0800
Subject: [PATCH] websocket 2pass bugfix
---
funasr/models/e2e_vad.py | 105 +++++++++++++++++++++++++++++++++++++++++++---------
1 files changed, 87 insertions(+), 18 deletions(-)
diff --git a/funasr/models/e2e_vad.py b/funasr/models/e2e_vad.py
old mode 100755
new mode 100644
index b64c677..82d8422
--- a/funasr/models/e2e_vad.py
+++ b/funasr/models/e2e_vad.py
@@ -35,6 +35,11 @@
class VADXOptions:
+ """
+ Author: Speech Lab of DAMO Academy, Alibaba Group
+ Deep-FSMN for Large Vocabulary Continuous Speech Recognition
+ https://arxiv.org/abs/1803.05030
+ """
def __init__(
self,
sample_rate: int = 16000,
@@ -99,6 +104,11 @@
class E2EVadSpeechBufWithDoa(object):
+ """
+ Author: Speech Lab of DAMO Academy, Alibaba Group
+ Deep-FSMN for Large Vocabulary Continuous Speech Recognition
+ https://arxiv.org/abs/1803.05030
+ """
def __init__(self):
self.start_ms = 0
self.end_ms = 0
@@ -117,6 +127,11 @@
class E2EVadFrameProb(object):
+ """
+ Author: Speech Lab of DAMO Academy, Alibaba Group
+ Deep-FSMN for Large Vocabulary Continuous Speech Recognition
+ https://arxiv.org/abs/1803.05030
+ """
def __init__(self):
self.noise_prob = 0.0
self.speech_prob = 0.0
@@ -126,6 +141,11 @@
class WindowDetector(object):
+ """
+ Author: Speech Lab of DAMO Academy, Alibaba Group
+ Deep-FSMN for Large Vocabulary Continuous Speech Recognition
+ https://arxiv.org/abs/1803.05030
+ """
def __init__(self, window_size_ms: int, sil_to_speech_time: int,
speech_to_sil_time: int, frame_size_ms: int):
self.window_size_ms = window_size_ms
@@ -192,7 +212,12 @@
class E2EVadModel(nn.Module):
- def __init__(self, encoder: FSMN, vad_post_args: Dict[str, Any]):
+ """
+ Author: Speech Lab of DAMO Academy, Alibaba Group
+ Deep-FSMN for Large Vocabulary Continuous Speech Recognition
+ https://arxiv.org/abs/1803.05030
+ """
+ def __init__(self, encoder: FSMN, vad_post_args: Dict[str, Any], frontend=None):
super(E2EVadModel, self).__init__()
self.vad_opts = VADXOptions(**vad_post_args)
self.windows_detector = WindowDetector(self.vad_opts.window_size_ms,
@@ -201,7 +226,7 @@
self.vad_opts.frame_in_ms)
self.encoder = encoder
# init variables
- self.is_final_send = False
+ self.is_final = False
self.data_buf_start_frame = 0
self.frm_cnt = 0
self.latest_confirmed_speech_frame = 0
@@ -215,6 +240,7 @@
self.sil_pdf_ids = self.vad_opts.sil_pdf_ids
self.noise_average_decibel = -100.0
self.pre_end_silence_detected = False
+ self.next_seg = True
self.output_data_buf = []
self.output_data_buf_offset = 0
@@ -228,10 +254,10 @@
self.data_buf_all = None
self.waveform = None
self.ResetDetection()
+ self.frontend = frontend
def AllResetDetection(self):
- self.encoder.cache_reset() # reset the in_cache in self.encoder for next query or next long sentence
- self.is_final_send = False
+ self.is_final = False
self.data_buf_start_frame = 0
self.frm_cnt = 0
self.latest_confirmed_speech_frame = 0
@@ -245,6 +271,7 @@
self.sil_pdf_ids = self.vad_opts.sil_pdf_ids
self.noise_average_decibel = -100.0
self.pre_end_silence_detected = False
+ self.next_seg = True
self.output_data_buf = []
self.output_data_buf_offset = 0
@@ -283,8 +310,8 @@
10 * math.log10((self.waveform[0][offset: offset + frame_sample_length]).square().sum() + \
0.000001))
- def ComputeScores(self, feats: torch.Tensor) -> None:
- scores = self.encoder(feats) # return B * T * D
+ def ComputeScores(self, feats: torch.Tensor, in_cache: Dict[str, torch.Tensor]) -> None:
+ scores = self.encoder(feats, in_cache).to('cpu') # return B * T * D
assert scores.shape[1] == feats.shape[1], "The shape between feats and scores does not match"
self.vad_opts.nn_eval_block_size = scores.shape[1]
self.frm_cnt += scores.shape[1] # count total frames
@@ -306,7 +333,7 @@
expected_sample_number = int(frm_cnt * self.vad_opts.sample_rate * self.vad_opts.frame_in_ms / 1000)
if last_frm_is_end_point:
extra_sample = max(0, int(self.vad_opts.frame_length_ms * self.vad_opts.sample_rate / 1000 - \
- self.vad_opts.sample_rate * self.vad_opts.frame_in_ms / 1000))
+ self.vad_opts.sample_rate * self.vad_opts.frame_in_ms / 1000))
expected_sample_number += int(extra_sample)
if end_point_is_sent_end:
expected_sample_number = max(expected_sample_number, len(self.data_buf))
@@ -443,11 +470,13 @@
return frame_state
- def forward(self, feats: torch.Tensor, waveform: torch.tensor, is_final_send: bool = False) -> List[List[List[int]]]:
+ def forward(self, feats: torch.Tensor, waveform: torch.tensor, in_cache: Dict[str, torch.Tensor] = dict(),
+ is_final: bool = False
+ ) -> Tuple[List[List[List[int]]], Dict[str, torch.Tensor]]:
self.waveform = waveform # compute decibel for each frame
self.ComputeDecibel()
- self.ComputeScores(feats)
- if not is_final_send:
+ self.ComputeScores(feats, in_cache)
+ if not is_final:
self.DetectCommonFrames()
else:
self.DetectLastFrames()
@@ -456,16 +485,56 @@
segment_batch = []
if len(self.output_data_buf) > 0:
for i in range(self.output_data_buf_offset, len(self.output_data_buf)):
- if self.output_data_buf[i].contain_seg_start_point and self.output_data_buf[
- i].contain_seg_end_point:
- segment = [self.output_data_buf[i].start_ms, self.output_data_buf[i].end_ms]
- segment_batch.append(segment)
- self.output_data_buf_offset += 1 # need update this parameter
+ if not is_final and (not self.output_data_buf[i].contain_seg_start_point or not self.output_data_buf[
+ i].contain_seg_end_point):
+ continue
+ segment = [self.output_data_buf[i].start_ms, self.output_data_buf[i].end_ms]
+ segment_batch.append(segment)
+ self.output_data_buf_offset += 1 # need update this parameter
if segment_batch:
segments.append(segment_batch)
- if is_final_send:
- self.AllResetDetection()
- return segments
+ if is_final:
+ # reset class variables and clear the dict for the next query
+ self.AllResetDetection()
+ return segments, in_cache
+
+ def forward_online(self, feats: torch.Tensor, waveform: torch.tensor, in_cache: Dict[str, torch.Tensor] = dict(),
+ is_final: bool = False, max_end_sil: int = 800
+ ) -> Tuple[List[List[List[int]]], Dict[str, torch.Tensor]]:
+ self.max_end_sil_frame_cnt_thresh = max_end_sil - self.vad_opts.speech_to_sil_time_thres
+ self.waveform = waveform # compute decibel for each frame
+
+ self.ComputeScores(feats, in_cache)
+ self.ComputeDecibel()
+ if not is_final:
+ self.DetectCommonFrames()
+ else:
+ self.DetectLastFrames()
+ segments = []
+ for batch_num in range(0, feats.shape[0]): # only support batch_size = 1 now
+ segment_batch = []
+ if len(self.output_data_buf) > 0:
+ for i in range(self.output_data_buf_offset, len(self.output_data_buf)):
+ if not self.output_data_buf[i].contain_seg_start_point:
+ continue
+ if not self.next_seg and not self.output_data_buf[i].contain_seg_end_point:
+ continue
+ start_ms = self.output_data_buf[i].start_ms if self.next_seg else -1
+ if self.output_data_buf[i].contain_seg_end_point:
+ end_ms = self.output_data_buf[i].end_ms
+ self.next_seg = True
+ self.output_data_buf_offset += 1
+ else:
+ end_ms = -1
+ self.next_seg = False
+ segment = [start_ms, end_ms]
+ segment_batch.append(segment)
+ if segment_batch:
+ segments.append(segment_batch)
+ if is_final:
+ # reset class variables and clear the dict for the next query
+ self.AllResetDetection()
+ return segments, in_cache
def DetectCommonFrames(self) -> int:
if self.vad_state_machine == VadStateMachine.kVadInStateEndPointDetected:
--
Gitblit v1.9.1