From c0b186b5b6e950472920964932ba3de546e06dbf Mon Sep 17 00:00:00 2001
From: 游雁 <zhifu.gzf@alibaba-inc.com>
Date: 星期五, 12 一月 2024 22:48:30 +0800
Subject: [PATCH] funasr1.0 streaming
---
funasr/models/fsmn_vad_streaming/model.py | 479 +++++++++++++++++++++++++++++------------------------------
1 files changed, 232 insertions(+), 247 deletions(-)
diff --git a/funasr/models/fsmn_vad_streaming/model.py b/funasr/models/fsmn_vad_streaming/model.py
index 9ceacf6..544fab8 100644
--- a/funasr/models/fsmn_vad_streaming/model.py
+++ b/funasr/models/fsmn_vad_streaming/model.py
@@ -11,7 +11,8 @@
from funasr.register import tables
from funasr.utils.load_utils import load_audio_text_image_video,extract_fbank
from funasr.utils.datadir_writer import DatadirWriter
-from torch.nn.utils.rnn import pad_sequence
+
+from dataclasses import dataclass
class VadStateMachine(Enum):
kVadInStateStartPointNotDetected = 1
@@ -38,7 +39,6 @@
class VadDetectMode(Enum):
kVadSingleUtteranceDetectMode = 0
kVadMutipleUtteranceDetectMode = 1
-
class VADXOptions:
"""
@@ -153,8 +153,10 @@
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):
+ 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
self.sil_to_speech_time = sil_to_speech_time
self.speech_to_sil_time = speech_to_sil_time
@@ -187,7 +189,7 @@
def GetWinSize(self) -> int:
return int(self.win_size_frame)
- def DetectOneFrame(self, frameState: FrameState, frame_count: int) -> AudioChangeState:
+ def DetectOneFrame(self, frameState: FrameState, frame_count: int, cache: dict={}) -> AudioChangeState:
cur_frame_state = FrameState.kFrameStateSil
if frameState == FrameState.kFrameStateSpeech:
cur_frame_state = 1
@@ -218,6 +220,38 @@
return int(self.frame_size_ms)
+@dataclass
+class StatsItem:
+
+ # init variables
+ data_buf_start_frame = 0
+ frm_cnt = 0
+ latest_confirmed_speech_frame = 0
+ lastest_confirmed_silence_frame = -1
+ continous_silence_frame_count = 0
+ vad_state_machine = VadStateMachine.kVadInStateStartPointNotDetected
+ confirmed_start_frame = -1
+ confirmed_end_frame = -1
+ number_end_time_detected = 0
+ sil_frame = 0
+ sil_pdf_ids: list
+ noise_average_decibel = -100.0
+ pre_end_silence_detected = False
+ next_seg = True # unused
+
+ output_data_buf = []
+ output_data_buf_offset = 0
+ frame_probs = [] # unused
+ max_end_sil_frame_cnt_thresh: int
+ speech_noise_thres: float
+ scores = None
+ max_time_out = False #unused
+ decibel = []
+ data_buf = None
+ data_buf_all = None
+ waveform = None
+ last_drop_frames = 0
+
@tables.register("model_classes", "FsmnVADStreaming")
class FsmnVADStreaming(nn.Module):
"""
@@ -233,143 +267,82 @@
):
super().__init__()
self.vad_opts = VADXOptions(**kwargs)
- self.windows_detector = WindowDetector(self.vad_opts.window_size_ms,
- self.vad_opts.sil_to_speech_time_thres,
- self.vad_opts.speech_to_sil_time_thres,
- self.vad_opts.frame_in_ms)
-
+
encoder_class = tables.encoder_classes.get(encoder.lower())
encoder = encoder_class(**encoder_conf)
self.encoder = encoder
- # init variables
- self.data_buf_start_frame = 0
- self.frm_cnt = 0
- self.latest_confirmed_speech_frame = 0
- self.lastest_confirmed_silence_frame = -1
- self.continous_silence_frame_count = 0
- self.vad_state_machine = VadStateMachine.kVadInStateStartPointNotDetected
- self.confirmed_start_frame = -1
- self.confirmed_end_frame = -1
- self.number_end_time_detected = 0
- self.sil_frame = 0
- 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
- self.frame_probs = []
- self.max_end_sil_frame_cnt_thresh = self.vad_opts.max_end_silence_time - self.vad_opts.speech_to_sil_time_thres
- self.speech_noise_thres = self.vad_opts.speech_noise_thres
- self.scores = None
- self.max_time_out = False
- self.decibel = []
- self.data_buf = None
- self.data_buf_all = None
- self.waveform = None
- self.last_drop_frames = 0
- def AllResetDetection(self):
- self.data_buf_start_frame = 0
- self.frm_cnt = 0
- self.latest_confirmed_speech_frame = 0
- self.lastest_confirmed_silence_frame = -1
- self.continous_silence_frame_count = 0
- self.vad_state_machine = VadStateMachine.kVadInStateStartPointNotDetected
- self.confirmed_start_frame = -1
- self.confirmed_end_frame = -1
- self.number_end_time_detected = 0
- self.sil_frame = 0
- 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
+ def ResetDetection(self, cache: dict = {}):
+ cache["stats"].continous_silence_frame_count = 0
+ cache["stats"].latest_confirmed_speech_frame = 0
+ cache["stats"].lastest_confirmed_silence_frame = -1
+ cache["stats"].confirmed_start_frame = -1
+ cache["stats"].confirmed_end_frame = -1
+ cache["stats"].vad_state_machine = VadStateMachine.kVadInStateStartPointNotDetected
+ cache["windows_detector"].Reset()
+ cache["stats"].sil_frame = 0
+ cache["stats"].frame_probs = []
- self.output_data_buf = []
- self.output_data_buf_offset = 0
- self.frame_probs = []
- self.max_end_sil_frame_cnt_thresh = self.vad_opts.max_end_silence_time - self.vad_opts.speech_to_sil_time_thres
- self.speech_noise_thres = self.vad_opts.speech_noise_thres
- self.scores = None
- self.max_time_out = False
- self.decibel = []
- self.data_buf = None
- self.data_buf_all = None
- self.waveform = None
- self.last_drop_frames = 0
- self.windows_detector.Reset()
+ if cache["stats"].output_data_buf:
+ assert cache["stats"].output_data_buf[-1].contain_seg_end_point == True
+ drop_frames = int(cache["stats"].output_data_buf[-1].end_ms / self.vad_opts.frame_in_ms)
+ real_drop_frames = drop_frames - cache["stats"].last_drop_frames
+ cache["stats"].last_drop_frames = drop_frames
+ cache["stats"].data_buf_all = cache["stats"].data_buf_all[real_drop_frames * int(self.vad_opts.frame_in_ms * self.vad_opts.sample_rate / 1000):]
+ cache["stats"].decibel = cache["stats"].decibel[real_drop_frames:]
+ cache["stats"].scores = cache["stats"].scores[:, real_drop_frames:, :]
- def ResetDetection(self):
- self.continous_silence_frame_count = 0
- self.latest_confirmed_speech_frame = 0
- self.lastest_confirmed_silence_frame = -1
- self.confirmed_start_frame = -1
- self.confirmed_end_frame = -1
- self.vad_state_machine = VadStateMachine.kVadInStateStartPointNotDetected
- self.windows_detector.Reset()
- self.sil_frame = 0
- self.frame_probs = []
-
- if self.output_data_buf:
- assert self.output_data_buf[-1].contain_seg_end_point == True
- drop_frames = int(self.output_data_buf[-1].end_ms / self.vad_opts.frame_in_ms)
- real_drop_frames = drop_frames - self.last_drop_frames
- self.last_drop_frames = drop_frames
- self.data_buf_all = self.data_buf_all[real_drop_frames * int(self.vad_opts.frame_in_ms * self.vad_opts.sample_rate / 1000):]
- self.decibel = self.decibel[real_drop_frames:]
- self.scores = self.scores[:, real_drop_frames:, :]
-
- def ComputeDecibel(self) -> None:
+ def ComputeDecibel(self, cache: dict = {}) -> None:
frame_sample_length = int(self.vad_opts.frame_length_ms * self.vad_opts.sample_rate / 1000)
frame_shift_length = int(self.vad_opts.frame_in_ms * self.vad_opts.sample_rate / 1000)
- if self.data_buf_all is None:
- self.data_buf_all = self.waveform[0] # self.data_buf is pointed to self.waveform[0]
- self.data_buf = self.data_buf_all
+ if cache["stats"].data_buf_all is None:
+ cache["stats"].data_buf_all = cache["stats"].waveform[0] # cache["stats"].data_buf is pointed to cache["stats"].waveform[0]
+ cache["stats"].data_buf = cache["stats"].data_buf_all
else:
- self.data_buf_all = torch.cat((self.data_buf_all, self.waveform[0]))
- for offset in range(0, self.waveform.shape[1] - frame_sample_length + 1, frame_shift_length):
- self.decibel.append(
- 10 * math.log10((self.waveform[0][offset: offset + frame_sample_length]).square().sum() + \
+ cache["stats"].data_buf_all = torch.cat((cache["stats"].data_buf_all, cache["stats"].waveform[0]))
+ for offset in range(0, cache["stats"].waveform.shape[1] - frame_sample_length + 1, frame_shift_length):
+ cache["stats"].decibel.append(
+ 10 * math.log10((cache["stats"].waveform[0][offset: offset + frame_sample_length]).square().sum() + \
0.000001))
- def ComputeScores(self, feats: torch.Tensor, cache: Dict[str, torch.Tensor]) -> None:
- scores = self.encoder(feats, cache).to('cpu') # return B * T * D
+ def ComputeScores(self, feats: torch.Tensor, cache: dict = {}) -> None:
+ scores = self.encoder(feats, cache=cache["encoder"]).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
- if self.scores is None:
- self.scores = scores # the first calculation
+ cache["stats"].frm_cnt += scores.shape[1] # count total frames
+ if cache["stats"].scores is None:
+ cache["stats"].scores = scores # the first calculation
else:
- self.scores = torch.cat((self.scores, scores), dim=1)
+ cache["stats"].scores = torch.cat((cache["stats"].scores, scores), dim=1)
- def PopDataBufTillFrame(self, frame_idx: int) -> None: # need check again
- while self.data_buf_start_frame < frame_idx:
- if len(self.data_buf) >= int(self.vad_opts.frame_in_ms * self.vad_opts.sample_rate / 1000):
- self.data_buf_start_frame += 1
- self.data_buf = self.data_buf_all[(self.data_buf_start_frame - self.last_drop_frames) * int(
+ def PopDataBufTillFrame(self, frame_idx: int, cache: dict={}) -> None: # need check again
+ while cache["stats"].data_buf_start_frame < frame_idx:
+ if len(cache["stats"].data_buf) >= int(self.vad_opts.frame_in_ms * self.vad_opts.sample_rate / 1000):
+ cache["stats"].data_buf_start_frame += 1
+ cache["stats"].data_buf = cache["stats"].data_buf_all[(cache["stats"].data_buf_start_frame - cache["stats"].last_drop_frames) * int(
self.vad_opts.frame_in_ms * self.vad_opts.sample_rate / 1000):]
def PopDataToOutputBuf(self, start_frm: int, frm_cnt: int, first_frm_is_start_point: bool,
- last_frm_is_end_point: bool, end_point_is_sent_end: bool) -> None:
- self.PopDataBufTillFrame(start_frm)
+ last_frm_is_end_point: bool, end_point_is_sent_end: bool, cache: dict={}) -> None:
+ self.PopDataBufTillFrame(start_frm, cache=cache)
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))
expected_sample_number += int(extra_sample)
if end_point_is_sent_end:
- expected_sample_number = max(expected_sample_number, len(self.data_buf))
- if len(self.data_buf) < expected_sample_number:
+ expected_sample_number = max(expected_sample_number, len(cache["stats"].data_buf))
+ if len(cache["stats"].data_buf) < expected_sample_number:
print('error in calling pop data_buf\n')
- if len(self.output_data_buf) == 0 or first_frm_is_start_point:
- self.output_data_buf.append(E2EVadSpeechBufWithDoa())
- self.output_data_buf[-1].Reset()
- self.output_data_buf[-1].start_ms = start_frm * self.vad_opts.frame_in_ms
- self.output_data_buf[-1].end_ms = self.output_data_buf[-1].start_ms
- self.output_data_buf[-1].doa = 0
- cur_seg = self.output_data_buf[-1]
+ if len(cache["stats"].output_data_buf) == 0 or first_frm_is_start_point:
+ cache["stats"].output_data_buf.append(E2EVadSpeechBufWithDoa())
+ cache["stats"].output_data_buf[-1].Reset()
+ cache["stats"].output_data_buf[-1].start_ms = start_frm * self.vad_opts.frame_in_ms
+ cache["stats"].output_data_buf[-1].end_ms = cache["stats"].output_data_buf[-1].start_ms
+ cache["stats"].output_data_buf[-1].doa = 0
+ cur_seg = cache["stats"].output_data_buf[-1]
if cur_seg.end_ms != start_frm * self.vad_opts.frame_in_ms:
print('warning\n')
out_pos = len(cur_seg.buffer) # cur_seg.buff鐜板湪娌″仛浠讳綍鎿嶄綔
@@ -378,10 +351,10 @@
data_to_pop = expected_sample_number
else:
data_to_pop = int(frm_cnt * self.vad_opts.frame_in_ms * self.vad_opts.sample_rate / 1000)
- if data_to_pop > len(self.data_buf):
- print('VAD data_to_pop is bigger than self.data_buf.size()!!!\n')
- data_to_pop = len(self.data_buf)
- expected_sample_number = len(self.data_buf)
+ if data_to_pop > len(cache["stats"].data_buf):
+ print('VAD data_to_pop is bigger than cache["stats"].data_buf.size()!!!\n')
+ data_to_pop = len(cache["stats"].data_buf)
+ expected_sample_number = len(cache["stats"].data_buf)
cur_seg.doa = 0
for sample_cpy_out in range(0, data_to_pop):
@@ -392,79 +365,79 @@
out_pos += 1
if cur_seg.end_ms != start_frm * self.vad_opts.frame_in_ms:
print('Something wrong with the VAD algorithm\n')
- self.data_buf_start_frame += frm_cnt
+ cache["stats"].data_buf_start_frame += frm_cnt
cur_seg.end_ms = (start_frm + frm_cnt) * self.vad_opts.frame_in_ms
if first_frm_is_start_point:
cur_seg.contain_seg_start_point = True
if last_frm_is_end_point:
cur_seg.contain_seg_end_point = True
- def OnSilenceDetected(self, valid_frame: int):
- self.lastest_confirmed_silence_frame = valid_frame
- if self.vad_state_machine == VadStateMachine.kVadInStateStartPointNotDetected:
- self.PopDataBufTillFrame(valid_frame)
+ def OnSilenceDetected(self, valid_frame: int, cache: dict = {}):
+ cache["stats"].lastest_confirmed_silence_frame = valid_frame
+ if cache["stats"].vad_state_machine == VadStateMachine.kVadInStateStartPointNotDetected:
+ self.PopDataBufTillFrame(valid_frame, cache=cache)
# silence_detected_callback_
# pass
- def OnVoiceDetected(self, valid_frame: int) -> None:
- self.latest_confirmed_speech_frame = valid_frame
- self.PopDataToOutputBuf(valid_frame, 1, False, False, False)
+ def OnVoiceDetected(self, valid_frame: int, cache:dict={}) -> None:
+ cache["stats"].latest_confirmed_speech_frame = valid_frame
+ self.PopDataToOutputBuf(valid_frame, 1, False, False, False, cache=cache)
- def OnVoiceStart(self, start_frame: int, fake_result: bool = False) -> None:
+ def OnVoiceStart(self, start_frame: int, fake_result: bool = False, cache:dict={}) -> None:
if self.vad_opts.do_start_point_detection:
pass
- if self.confirmed_start_frame != -1:
+ if cache["stats"].confirmed_start_frame != -1:
print('not reset vad properly\n')
else:
- self.confirmed_start_frame = start_frame
+ cache["stats"].confirmed_start_frame = start_frame
- if not fake_result and self.vad_state_machine == VadStateMachine.kVadInStateStartPointNotDetected:
- self.PopDataToOutputBuf(self.confirmed_start_frame, 1, True, False, False)
+ if not fake_result and cache["stats"].vad_state_machine == VadStateMachine.kVadInStateStartPointNotDetected:
+ self.PopDataToOutputBuf(cache["stats"].confirmed_start_frame, 1, True, False, False, cache=cache)
- def OnVoiceEnd(self, end_frame: int, fake_result: bool, is_last_frame: bool) -> None:
- for t in range(self.latest_confirmed_speech_frame + 1, end_frame):
- self.OnVoiceDetected(t)
+ def OnVoiceEnd(self, end_frame: int, fake_result: bool, is_last_frame: bool, cache:dict={}) -> None:
+ for t in range(cache["stats"].latest_confirmed_speech_frame + 1, end_frame):
+ self.OnVoiceDetected(t, cache=cache)
if self.vad_opts.do_end_point_detection:
pass
- if self.confirmed_end_frame != -1:
+ if cache["stats"].confirmed_end_frame != -1:
print('not reset vad properly\n')
else:
- self.confirmed_end_frame = end_frame
+ cache["stats"].confirmed_end_frame = end_frame
if not fake_result:
- self.sil_frame = 0
- self.PopDataToOutputBuf(self.confirmed_end_frame, 1, False, True, is_last_frame)
- self.number_end_time_detected += 1
+ cache["stats"].sil_frame = 0
+ self.PopDataToOutputBuf(cache["stats"].confirmed_end_frame, 1, False, True, is_last_frame, cache=cache)
+ cache["stats"].number_end_time_detected += 1
- def MaybeOnVoiceEndIfLastFrame(self, is_final_frame: bool, cur_frm_idx: int) -> None:
+ def MaybeOnVoiceEndIfLastFrame(self, is_final_frame: bool, cur_frm_idx: int, cache: dict = {}) -> None:
if is_final_frame:
- self.OnVoiceEnd(cur_frm_idx, False, True)
- self.vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
+ self.OnVoiceEnd(cur_frm_idx, False, True, cache=cache)
+ cache["stats"].vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
- def GetLatency(self) -> int:
- return int(self.LatencyFrmNumAtStartPoint() * self.vad_opts.frame_in_ms)
+ def GetLatency(self, cache: dict = {}) -> int:
+ return int(self.LatencyFrmNumAtStartPoint(cache=cache) * self.vad_opts.frame_in_ms)
- def LatencyFrmNumAtStartPoint(self) -> int:
- vad_latency = self.windows_detector.GetWinSize()
+ def LatencyFrmNumAtStartPoint(self, cache: dict = {}) -> int:
+ vad_latency = cache["windows_detector"].GetWinSize()
if self.vad_opts.do_extend:
vad_latency += int(self.vad_opts.lookback_time_start_point / self.vad_opts.frame_in_ms)
return vad_latency
- def GetFrameState(self, t: int):
+ def GetFrameState(self, t: int, cache: dict = {}):
frame_state = FrameState.kFrameStateInvalid
- cur_decibel = self.decibel[t]
- cur_snr = cur_decibel - self.noise_average_decibel
+ cur_decibel = cache["stats"].decibel[t]
+ cur_snr = cur_decibel - cache["stats"].noise_average_decibel
# for each frame, calc log posterior probability of each state
if cur_decibel < self.vad_opts.decibel_thres:
frame_state = FrameState.kFrameStateSil
- self.DetectOneFrame(frame_state, t, False)
+ self.DetectOneFrame(frame_state, t, False, cache=cache)
return frame_state
sum_score = 0.0
noise_prob = 0.0
- assert len(self.sil_pdf_ids) == self.vad_opts.silence_pdf_num
- if len(self.sil_pdf_ids) > 0:
- assert len(self.scores) == 1 # 鍙敮鎸乥atch_size = 1鐨勬祴璇�
- sil_pdf_scores = [self.scores[0][t][sil_pdf_id] for sil_pdf_id in self.sil_pdf_ids]
+ assert len(cache["stats"].sil_pdf_ids) == self.vad_opts.silence_pdf_num
+ if len(cache["stats"].sil_pdf_ids) > 0:
+ assert len(cache["stats"].scores) == 1 # 鍙敮鎸乥atch_size = 1鐨勬祴璇�
+ sil_pdf_scores = [cache["stats"].scores[0][t][sil_pdf_id] for sil_pdf_id in cache["stats"].sil_pdf_ids]
sum_score = sum(sil_pdf_scores)
noise_prob = math.log(sum_score) * self.vad_opts.speech_2_noise_ratio
total_score = 1.0
@@ -476,58 +449,69 @@
frame_prob.speech_prob = speech_prob
frame_prob.score = sum_score
frame_prob.frame_id = t
- self.frame_probs.append(frame_prob)
- if math.exp(speech_prob) >= math.exp(noise_prob) + self.speech_noise_thres:
+ cache["stats"].frame_probs.append(frame_prob)
+ if math.exp(speech_prob) >= math.exp(noise_prob) + cache["stats"].speech_noise_thres:
if cur_snr >= self.vad_opts.snr_thres and cur_decibel >= self.vad_opts.decibel_thres:
frame_state = FrameState.kFrameStateSpeech
else:
frame_state = FrameState.kFrameStateSil
else:
frame_state = FrameState.kFrameStateSil
- if self.noise_average_decibel < -99.9:
- self.noise_average_decibel = cur_decibel
+ if cache["stats"].noise_average_decibel < -99.9:
+ cache["stats"].noise_average_decibel = cur_decibel
else:
- self.noise_average_decibel = (cur_decibel + self.noise_average_decibel * (
+ cache["stats"].noise_average_decibel = (cur_decibel + cache["stats"].noise_average_decibel * (
self.vad_opts.noise_frame_num_used_for_snr
- 1)) / self.vad_opts.noise_frame_num_used_for_snr
return frame_state
- def forward(self, feats: torch.Tensor, waveform: torch.tensor, cache: Dict[str, torch.Tensor] = dict(),
+ def forward(self, feats: torch.Tensor, waveform: torch.tensor, cache: dict = {},
is_final: bool = False
):
- if len(cache) == 0:
- self.AllResetDetection()
- self.waveform = waveform # compute decibel for each frame
- self.ComputeDecibel()
- self.ComputeScores(feats, cache)
+ # if len(cache) == 0:
+ # self.AllResetDetection()
+ # self.waveform = waveform # compute decibel for each frame
+ cache["stats"].waveform = waveform
+ self.ComputeDecibel(cache=cache)
+ self.ComputeScores(feats, cache=cache)
if not is_final:
- self.DetectCommonFrames()
+ self.DetectCommonFrames(cache=cache)
else:
- self.DetectLastFrames()
+ self.DetectLastFrames(cache=cache)
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 is_final and (not self.output_data_buf[i].contain_seg_start_point or not self.output_data_buf[
+ if len(cache["stats"].output_data_buf) > 0:
+ for i in range(cache["stats"].output_data_buf_offset, len(cache["stats"].output_data_buf)):
+ if not is_final and (not cache["stats"].output_data_buf[i].contain_seg_start_point or not cache["stats"].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 = [cache["stats"].output_data_buf[i].start_ms, cache["stats"].output_data_buf[i].end_ms]
segment_batch.append(segment)
- self.output_data_buf_offset += 1 # need update this parameter
+ cache["stats"].output_data_buf_offset += 1 # need update this parameter
if segment_batch:
segments.append(segment_batch)
- if is_final:
- # reset class variables and clear the dict for the next query
- self.AllResetDetection()
+ # if is_final:
+ # # reset class variables and clear the dict for the next query
+ # self.AllResetDetection()
return segments
def init_cache(self, cache: dict = {}, **kwargs):
cache["frontend"] = {}
cache["prev_samples"] = torch.empty(0)
cache["encoder"] = {}
-
+ windows_detector = WindowDetector(self.vad_opts.window_size_ms,
+ self.vad_opts.sil_to_speech_time_thres,
+ self.vad_opts.speech_to_sil_time_thres,
+ self.vad_opts.frame_in_ms)
+
+ stats = StatsItem(sil_pdf_ids=self.vad_opts.sil_pdf_ids,
+ max_end_sil_frame_cnt_thresh=self.vad_opts.max_end_silence_time - self.vad_opts.speech_to_sil_time_thres,
+ speech_noise_thres=self.vad_opts.speech_noise_thres,
+ )
+ cache["windows_detector"] = windows_detector
+ cache["stats"] = stats
return cache
def generate(self,
@@ -544,7 +528,7 @@
self.init_cache(cache, **kwargs)
meta_data = {}
- chunk_size = kwargs.get("chunk_size", 50) # 50ms
+ chunk_size = kwargs.get("chunk_size", 60000) # 50ms
chunk_stride_samples = int(chunk_size * frontend.fs / 1000)
time1 = time.perf_counter()
@@ -585,10 +569,11 @@
"feats": speech,
"waveform": cache["frontend"]["waveforms"],
"is_final": kwargs["is_final"],
- "cache": cache["encoder"]
+ "cache": cache
}
segments_i = self.forward(**batch)
- segments.extend(segments_i)
+ if len(segments_i) > 0:
+ segments.extend(*segments_i)
cache["prev_samples"] = audio_sample[:-m]
@@ -614,30 +599,30 @@
return results, meta_data
- def DetectCommonFrames(self) -> int:
- if self.vad_state_machine == VadStateMachine.kVadInStateEndPointDetected:
+ def DetectCommonFrames(self, cache: dict = {}) -> int:
+ if cache["stats"].vad_state_machine == VadStateMachine.kVadInStateEndPointDetected:
return 0
for i in range(self.vad_opts.nn_eval_block_size - 1, -1, -1):
frame_state = FrameState.kFrameStateInvalid
- frame_state = self.GetFrameState(self.frm_cnt - 1 - i - self.last_drop_frames)
- self.DetectOneFrame(frame_state, self.frm_cnt - 1 - i, False)
+ frame_state = self.GetFrameState(cache["stats"].frm_cnt - 1 - i - cache["stats"].last_drop_frames, cache=cache)
+ self.DetectOneFrame(frame_state, cache["stats"].frm_cnt - 1 - i, False, cache=cache)
return 0
- def DetectLastFrames(self) -> int:
- if self.vad_state_machine == VadStateMachine.kVadInStateEndPointDetected:
+ def DetectLastFrames(self, cache: dict = {}) -> int:
+ if cache["stats"].vad_state_machine == VadStateMachine.kVadInStateEndPointDetected:
return 0
for i in range(self.vad_opts.nn_eval_block_size - 1, -1, -1):
frame_state = FrameState.kFrameStateInvalid
- frame_state = self.GetFrameState(self.frm_cnt - 1 - i - self.last_drop_frames)
+ frame_state = self.GetFrameState(cache["stats"].frm_cnt - 1 - i - cache["stats"].last_drop_frames, cache=cache)
if i != 0:
- self.DetectOneFrame(frame_state, self.frm_cnt - 1 - i, False)
+ self.DetectOneFrame(frame_state, cache["stats"].frm_cnt - 1 - i, False, cache=cache)
else:
- self.DetectOneFrame(frame_state, self.frm_cnt - 1, True)
+ self.DetectOneFrame(frame_state, cache["stats"].frm_cnt - 1, True, cache=cache)
return 0
- def DetectOneFrame(self, cur_frm_state: FrameState, cur_frm_idx: int, is_final_frame: bool) -> None:
+ def DetectOneFrame(self, cur_frm_state: FrameState, cur_frm_idx: int, is_final_frame: bool, cache: dict = {}) -> None:
tmp_cur_frm_state = FrameState.kFrameStateInvalid
if cur_frm_state == FrameState.kFrameStateSpeech:
if math.fabs(1.0) > self.vad_opts.fe_prior_thres:
@@ -646,101 +631,101 @@
tmp_cur_frm_state = FrameState.kFrameStateSil
elif cur_frm_state == FrameState.kFrameStateSil:
tmp_cur_frm_state = FrameState.kFrameStateSil
- state_change = self.windows_detector.DetectOneFrame(tmp_cur_frm_state, cur_frm_idx)
+ state_change = cache["windows_detector"].DetectOneFrame(tmp_cur_frm_state, cur_frm_idx, cache=cache)
frm_shift_in_ms = self.vad_opts.frame_in_ms
if AudioChangeState.kChangeStateSil2Speech == state_change:
- silence_frame_count = self.continous_silence_frame_count
- self.continous_silence_frame_count = 0
- self.pre_end_silence_detected = False
+ silence_frame_count = cache["stats"].continous_silence_frame_count
+ cache["stats"].continous_silence_frame_count = 0
+ cache["stats"].pre_end_silence_detected = False
start_frame = 0
- if self.vad_state_machine == VadStateMachine.kVadInStateStartPointNotDetected:
- start_frame = max(self.data_buf_start_frame, cur_frm_idx - self.LatencyFrmNumAtStartPoint())
- self.OnVoiceStart(start_frame)
- self.vad_state_machine = VadStateMachine.kVadInStateInSpeechSegment
+ if cache["stats"].vad_state_machine == VadStateMachine.kVadInStateStartPointNotDetected:
+ start_frame = max(cache["stats"].data_buf_start_frame, cur_frm_idx - self.LatencyFrmNumAtStartPoint(cache=cache))
+ self.OnVoiceStart(start_frame, cache=cache)
+ cache["stats"].vad_state_machine = VadStateMachine.kVadInStateInSpeechSegment
for t in range(start_frame + 1, cur_frm_idx + 1):
- self.OnVoiceDetected(t)
- elif self.vad_state_machine == VadStateMachine.kVadInStateInSpeechSegment:
- for t in range(self.latest_confirmed_speech_frame + 1, cur_frm_idx):
- self.OnVoiceDetected(t)
- if cur_frm_idx - self.confirmed_start_frame + 1 > \
+ self.OnVoiceDetected(t, cache=cache)
+ elif cache["stats"].vad_state_machine == VadStateMachine.kVadInStateInSpeechSegment:
+ for t in range(cache["stats"].latest_confirmed_speech_frame + 1, cur_frm_idx):
+ self.OnVoiceDetected(t, cache=cache)
+ if cur_frm_idx - cache["stats"].confirmed_start_frame + 1 > \
self.vad_opts.max_single_segment_time / frm_shift_in_ms:
- self.OnVoiceEnd(cur_frm_idx, False, False)
- self.vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
+ self.OnVoiceEnd(cur_frm_idx, False, False, cache=cache)
+ cache["stats"].vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
elif not is_final_frame:
- self.OnVoiceDetected(cur_frm_idx)
+ self.OnVoiceDetected(cur_frm_idx, cache=cache)
else:
- self.MaybeOnVoiceEndIfLastFrame(is_final_frame, cur_frm_idx)
+ self.MaybeOnVoiceEndIfLastFrame(is_final_frame, cur_frm_idx, cache=cache)
else:
pass
elif AudioChangeState.kChangeStateSpeech2Sil == state_change:
- self.continous_silence_frame_count = 0
- if self.vad_state_machine == VadStateMachine.kVadInStateStartPointNotDetected:
+ cache["stats"].continous_silence_frame_count = 0
+ if cache["stats"].vad_state_machine == VadStateMachine.kVadInStateStartPointNotDetected:
pass
- elif self.vad_state_machine == VadStateMachine.kVadInStateInSpeechSegment:
- if cur_frm_idx - self.confirmed_start_frame + 1 > \
+ elif cache["stats"].vad_state_machine == VadStateMachine.kVadInStateInSpeechSegment:
+ if cur_frm_idx - cache["stats"].confirmed_start_frame + 1 > \
self.vad_opts.max_single_segment_time / frm_shift_in_ms:
- self.OnVoiceEnd(cur_frm_idx, False, False)
- self.vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
+ self.OnVoiceEnd(cur_frm_idx, False, False, cache=cache)
+ cache["stats"].vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
elif not is_final_frame:
- self.OnVoiceDetected(cur_frm_idx)
+ self.OnVoiceDetected(cur_frm_idx, cache=cache)
else:
- self.MaybeOnVoiceEndIfLastFrame(is_final_frame, cur_frm_idx)
+ self.MaybeOnVoiceEndIfLastFrame(is_final_frame, cur_frm_idx, cache=cache)
else:
pass
elif AudioChangeState.kChangeStateSpeech2Speech == state_change:
- self.continous_silence_frame_count = 0
- if self.vad_state_machine == VadStateMachine.kVadInStateInSpeechSegment:
- if cur_frm_idx - self.confirmed_start_frame + 1 > \
+ cache["stats"].continous_silence_frame_count = 0
+ if cache["stats"].vad_state_machine == VadStateMachine.kVadInStateInSpeechSegment:
+ if cur_frm_idx - cache["stats"].confirmed_start_frame + 1 > \
self.vad_opts.max_single_segment_time / frm_shift_in_ms:
- self.max_time_out = True
- self.OnVoiceEnd(cur_frm_idx, False, False)
- self.vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
+ cache["stats"].max_time_out = True
+ self.OnVoiceEnd(cur_frm_idx, False, False, cache=cache)
+ cache["stats"].vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
elif not is_final_frame:
- self.OnVoiceDetected(cur_frm_idx)
+ self.OnVoiceDetected(cur_frm_idx, cache=cache)
else:
- self.MaybeOnVoiceEndIfLastFrame(is_final_frame, cur_frm_idx)
+ self.MaybeOnVoiceEndIfLastFrame(is_final_frame, cur_frm_idx, cache=cache)
else:
pass
elif AudioChangeState.kChangeStateSil2Sil == state_change:
- self.continous_silence_frame_count += 1
- if self.vad_state_machine == VadStateMachine.kVadInStateStartPointNotDetected:
+ cache["stats"].continous_silence_frame_count += 1
+ if cache["stats"].vad_state_machine == VadStateMachine.kVadInStateStartPointNotDetected:
# silence timeout, return zero length decision
if ((self.vad_opts.detect_mode == VadDetectMode.kVadSingleUtteranceDetectMode.value) and (
- self.continous_silence_frame_count * frm_shift_in_ms > self.vad_opts.max_start_silence_time)) \
- or (is_final_frame and self.number_end_time_detected == 0):
- for t in range(self.lastest_confirmed_silence_frame + 1, cur_frm_idx):
- self.OnSilenceDetected(t)
- self.OnVoiceStart(0, True)
- self.OnVoiceEnd(0, True, False);
- self.vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
+ cache["stats"].continous_silence_frame_count * frm_shift_in_ms > self.vad_opts.max_start_silence_time)) \
+ or (is_final_frame and cache["stats"].number_end_time_detected == 0):
+ for t in range(cache["stats"].lastest_confirmed_silence_frame + 1, cur_frm_idx):
+ self.OnSilenceDetected(t, cache=cache)
+ self.OnVoiceStart(0, True, cache=cache)
+ self.OnVoiceEnd(0, True, False, cache=cache)
+ cache["stats"].vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
else:
- if cur_frm_idx >= self.LatencyFrmNumAtStartPoint():
- self.OnSilenceDetected(cur_frm_idx - self.LatencyFrmNumAtStartPoint())
- elif self.vad_state_machine == VadStateMachine.kVadInStateInSpeechSegment:
- if self.continous_silence_frame_count * frm_shift_in_ms >= self.max_end_sil_frame_cnt_thresh:
- lookback_frame = int(self.max_end_sil_frame_cnt_thresh / frm_shift_in_ms)
+ if cur_frm_idx >= self.LatencyFrmNumAtStartPoint(cache=cache):
+ self.OnSilenceDetected(cur_frm_idx - self.LatencyFrmNumAtStartPoint(cache=cache), cache=cache)
+ elif cache["stats"].vad_state_machine == VadStateMachine.kVadInStateInSpeechSegment:
+ if cache["stats"].continous_silence_frame_count * frm_shift_in_ms >= cache["stats"].max_end_sil_frame_cnt_thresh:
+ lookback_frame = int(cache["stats"].max_end_sil_frame_cnt_thresh / frm_shift_in_ms)
if self.vad_opts.do_extend:
lookback_frame -= int(self.vad_opts.lookahead_time_end_point / frm_shift_in_ms)
lookback_frame -= 1
lookback_frame = max(0, lookback_frame)
- self.OnVoiceEnd(cur_frm_idx - lookback_frame, False, False)
- self.vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
- elif cur_frm_idx - self.confirmed_start_frame + 1 > \
+ self.OnVoiceEnd(cur_frm_idx - lookback_frame, False, False, cache=cache)
+ cache["stats"].vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
+ elif cur_frm_idx - cache["stats"].confirmed_start_frame + 1 > \
self.vad_opts.max_single_segment_time / frm_shift_in_ms:
- self.OnVoiceEnd(cur_frm_idx, False, False)
- self.vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
+ self.OnVoiceEnd(cur_frm_idx, False, False, cache=cache)
+ cache["stats"].vad_state_machine = VadStateMachine.kVadInStateEndPointDetected
elif self.vad_opts.do_extend and not is_final_frame:
- if self.continous_silence_frame_count <= int(
+ if cache["stats"].continous_silence_frame_count <= int(
self.vad_opts.lookahead_time_end_point / frm_shift_in_ms):
- self.OnVoiceDetected(cur_frm_idx)
+ self.OnVoiceDetected(cur_frm_idx, cache=cache)
else:
- self.MaybeOnVoiceEndIfLastFrame(is_final_frame, cur_frm_idx)
+ self.MaybeOnVoiceEndIfLastFrame(is_final_frame, cur_frm_idx, cache=cache)
else:
pass
- if self.vad_state_machine == VadStateMachine.kVadInStateEndPointDetected and \
+ if cache["stats"].vad_state_machine == VadStateMachine.kVadInStateEndPointDetected and \
self.vad_opts.detect_mode == VadDetectMode.kVadMutipleUtteranceDetectMode.value:
- self.ResetDetection()
+ self.ResetDetection(cache=cache)
--
Gitblit v1.9.1