前置唤醒词识别优化
This commit is contained in:
parent
14c95ab079
commit
84551d59eb
136
core/recorder.py
136
core/recorder.py
@ -16,11 +16,68 @@ import tempfile
|
||||
import wave
|
||||
from core import fay_core
|
||||
from core import interact
|
||||
# ===== 新增：用于前置唤醒词句首容错 =====
|
||||
import re
|
||||
import unicodedata
|
||||
|
||||
# 启动时间 (秒)
|
||||
_ATTACK = 0.2
|
||||
_ATTACK = 0.08 # ↓ 改小：让系统更早进入拾音，避免“唤醒词前半截被吃掉”
|
||||
|
||||
# 释放时间 (秒)
|
||||
_RELEASE = 0.7
|
||||
_RELEASE = 0.55 # ↓ 略微缩短，避免一句话被切成两段
|
||||
|
||||
# ===== 新增：前置唤醒词句首规范化与匹配 =====
|
||||
_PUNCS = "，。！？!?,.、:：；;“”\"'()（）[]【】<>《》-—…" # 常见中文标点
|
||||
_FILLER_PREFIX = ("嗯", "啊", "呃", "欸", "诶", "喂", "那个", "就是", "然后") # 常见句首语气词（ASR 很容易加）
|
||||
|
||||
def _norm_head(s: str) -> str:
|
||||
"""只做句首容错：去不可见/空白/句首标点/句首语气词，不改变正文结构。"""
|
||||
if not s:
|
||||
return ""
|
||||
s = unicodedata.normalize("NFKC", s).strip()
|
||||
# 去掉开头空白
|
||||
s = re.sub(r"^\s+", "", s)
|
||||
# 去掉开头标点（可重复）
|
||||
s = re.sub(r"^[{}]+".format(re.escape(_PUNCS)), "", s)
|
||||
|
||||
# 去掉句首常见语气词（允许多次叠加）
|
||||
changed = True
|
||||
while changed:
|
||||
changed = False
|
||||
for fp in _FILLER_PREFIX:
|
||||
if s.startswith(fp):
|
||||
s = s[len(fp):]
|
||||
s = re.sub(r"^\s+", "", s)
|
||||
s = re.sub(r"^[{}]+".format(re.escape(_PUNCS)), "", s)
|
||||
changed = True
|
||||
break
|
||||
return s
|
||||
|
||||
def _front_wake_match(text: str, wake_words):
|
||||
"""
|
||||
前置唤醒词匹配（严格前置）：
|
||||
- 唤醒词必须在规范化后的最前面
|
||||
- 不允许句中唤醒
|
||||
"""
|
||||
t = _norm_head(text)
|
||||
|
||||
for w in wake_words:
|
||||
w = w.strip()
|
||||
if not w:
|
||||
continue
|
||||
|
||||
# 允许：唤醒词后面紧跟空格/标点/语气助词
|
||||
# 例："小橄榄，帮我..." "小橄榄啊 帮我..."
|
||||
if t.startswith(w):
|
||||
rest = t[len(w):] # 去掉唤醒词，得到真正的问题
|
||||
# 去掉紧随其后的标点 / 空格 / 语气助词
|
||||
rest = rest.lstrip(" \t\r\n" + _PUNCS)
|
||||
rest = re.sub(r"^(啊|呀|呢|吧|哈|哎|诶|欸)\s*", "", rest)
|
||||
rest = rest.lstrip(" \t\r\n" + _PUNCS)
|
||||
return True, w, rest
|
||||
|
||||
return False, None, ""
|
||||
|
||||
|
||||
|
||||
class Recorder:
|
||||
@ -141,35 +198,45 @@ class Recorder:
|
||||
self.timer.cancel() # 取消之前的计时器任务
|
||||
self.timer = threading.Timer(60, self.reset_wakeup_status) # 重设计时器为60秒
|
||||
self.timer.start()
|
||||
|
||||
#前置唤醒词模式
|
||||
elif cfg.config['source']['wake_word_type'] == 'front':
|
||||
wake_word = cfg.config['source']['wake_word']
|
||||
wake_word_list = wake_word.split(',')
|
||||
wake_up = False
|
||||
for word in wake_word_list:
|
||||
if text.startswith(word):
|
||||
wake_up_word = word
|
||||
wake_up = True
|
||||
break
|
||||
if wake_up:
|
||||
|
||||
# 前置唤醒词模式（严格前置，但句首做容错）
|
||||
elif cfg.config['source']['wake_word_type'] == 'front':
|
||||
# 读取配置的唤醒词（支持多个）
|
||||
wake_word = cfg.config['source']['wake_word']
|
||||
wake_word_list = [w.strip() for w in wake_word.split(',') if w.strip()]
|
||||
|
||||
matched, wake_up_word, question = _front_wake_match(text, wake_word_list)
|
||||
|
||||
if matched:
|
||||
util.printInfo(1, self.username, "唤醒成功！")
|
||||
if wsa_server.get_web_instance().is_connected(self.username):
|
||||
wsa_server.get_web_instance().add_cmd({"panelMsg": "唤醒成功！", "Username" : self.username , 'robot': f'http://{cfg.fay_url}:5000/robot/Listening.jpg'})
|
||||
wsa_server.get_web_instance().add_cmd({"panelMsg": "唤醒成功！", "Username": self.username,
|
||||
'robot': f'http://{cfg.fay_url}:5000/robot/Listening.jpg'})
|
||||
if wsa_server.get_instance().is_connected(self.username):
|
||||
content = {'Topic': 'Unreal', 'Data': {'Key': 'log', 'Value': "唤醒成功！"}, 'Username' : self.username, 'robot': f'http://{cfg.fay_url}:5000/robot/Listening.jpg'}
|
||||
content = {'Topic': 'Unreal', 'Data': {'Key': 'log', 'Value': "唤醒成功！"},
|
||||
'Username': self.username,
|
||||
'robot': f'http://{cfg.fay_url}:5000/robot/Listening.jpg'}
|
||||
wsa_server.get_instance().add_cmd(content)
|
||||
#去除唤醒词后语句
|
||||
question = text#[len(wake_up_word):].lstrip()
|
||||
self.on_speaking(question)
|
||||
|
||||
# 在识别到【前置唤醒词】后，发送“去掉唤醒词后的问题”
|
||||
if question:
|
||||
self.on_speaking(question)
|
||||
else:
|
||||
intt = interact.Interact("auto_play", 2, {'user': self.username, 'text': "在呢，你说？"})
|
||||
self.__fay.on_interact(intt)
|
||||
|
||||
self.processing = False
|
||||
else:
|
||||
util.printInfo(1, self.username, "[!] 待唤醒！")
|
||||
if wsa_server.get_web_instance().is_connected(self.username):
|
||||
wsa_server.get_web_instance().add_cmd({"panelMsg": "[!] 待唤醒！", "Username" : self.username , 'robot': f'http://{cfg.fay_url}:5000/robot/Normal.jpg'})
|
||||
wsa_server.get_web_instance().add_cmd({"panelMsg": "[!] 待唤醒！", "Username": self.username,
|
||||
'robot': f'http://{cfg.fay_url}:5000/robot/Normal.jpg'})
|
||||
if wsa_server.get_instance().is_connected(self.username):
|
||||
content = {'Topic': 'Unreal', 'Data': {'Key': 'log', 'Value': "[!] 待唤醒！"}, 'Username' : self.username, 'robot': f'http://{cfg.fay_url}:5000/robot/Normal.jpg'}
|
||||
content = {'Topic': 'Unreal', 'Data': {'Key': 'log', 'Value': "[!] 待唤醒！"},
|
||||
'Username': self.username,
|
||||
'robot': f'http://{cfg.fay_url}:5000/robot/Normal.jpg'}
|
||||
wsa_server.get_instance().add_cmd(content)
|
||||
self.processing = False
|
||||
|
||||
#非唤醒模式
|
||||
else:
|
||||
@ -220,12 +287,8 @@ class Recorder:
|
||||
continue
|
||||
#是否可以拾音,不可以就掉弃录音
|
||||
can_listen = True
|
||||
#没有开唤醒，但面板或数字人正在播音时不能拾音
|
||||
if cfg.config['source']['wake_word_enabled'] == False and self.__fay.speaking == True:
|
||||
can_listen = False
|
||||
|
||||
#普通唤醒模式已经激活，并且面板或数字人正在输出声音时不能拾音
|
||||
if cfg.config['source']['wake_word_enabled'] == True and cfg.config['source']['wake_word_type'] == 'common' and self.wakeup_matched == True and self.__fay.speaking == True:
|
||||
if self.__fay.speaking == True:
|
||||
# 只要数字人/面板在播放TTS，就禁拾音，避免把自己的声音识别成用户输入
|
||||
can_listen = False
|
||||
|
||||
if can_listen == False:#掉弃录音
|
||||
@ -234,7 +297,7 @@ class Recorder:
|
||||
|
||||
#计算音量是否满足激活拾音
|
||||
level = audioop.rms(data, 2)
|
||||
if len(self.__history_data) >= 10:#保存激活前的音频，以免信息掉失
|
||||
if len(self.__history_data) >= 20:#保存激活前的音频，以免信息掉失
|
||||
self.__history_data.pop(0)
|
||||
if len(self.__history_level) >= 500:
|
||||
self.__history_level.pop(0)
|
||||
@ -242,12 +305,19 @@ class Recorder:
|
||||
self.__history_level.append(level)
|
||||
percentage = level / self.__MAX_LEVEL
|
||||
history_percentage = self.__get_history_percentage(30)
|
||||
|
||||
# ===== 改进：阈值平滑变化，避免断句导致唤醒词被截断 =====
|
||||
up_alpha = 0.01 # 环境变吵：慢慢升
|
||||
down_alpha = 0.05 # 环境变安静：也不要瞬间掉
|
||||
|
||||
if history_percentage > self.__dynamic_threshold:
|
||||
self.__dynamic_threshold += (history_percentage - self.__dynamic_threshold) * 0.0025
|
||||
elif history_percentage < self.__dynamic_threshold:
|
||||
self.__dynamic_threshold += (history_percentage - self.__dynamic_threshold) * 1
|
||||
|
||||
|
||||
self.__dynamic_threshold += (history_percentage - self.__dynamic_threshold) * up_alpha
|
||||
else:
|
||||
self.__dynamic_threshold += (history_percentage - self.__dynamic_threshold) * down_alpha
|
||||
|
||||
# 给阈值一个下限，防止过度灵敏
|
||||
self.__dynamic_threshold = max(self.__dynamic_threshold, 0.02)
|
||||
|
||||
#激活拾音
|
||||
if percentage > self.__dynamic_threshold:
|
||||
last_speaking_time = time.time()
|
||||
|
||||
Loading…
Reference in New Issue
Block a user