config.py

import pathlib
import re
import logging

DEBUG = False
LOG_LEVEL = logging.DEBUG if DEBUG else logging.INFO

logging.getLogger("pywhispercpp").setLevel(logging.WARNING)
logging.basicConfig(
    level=LOG_LEVEL,
    format="%(asctime)s - %(levelname)s - %(message)s",
    filename='translator.log',
    datefmt="%H:%M:%S"
)
# save pipelines data to disk
SAVE_DATA_SAVE = False
# Add terminal log
console_handler = logging.StreamHandler()
console_handler.setLevel(LOG_LEVEL)
console_formatter = logging.Formatter("%(asctime)s - %(levelname)s - %(message)s")
console_handler.setFormatter(console_formatter)
logging.getLogger().addHandler(console_handler)

# 文字输出长度阈值
TEXT_THREHOLD = 6
# 音频段的决策时间
DESIGN_TIME_THREHOLD = 3
# 最长语音时长
MAX_SPEECH_DURATION_S = 15

BASE_DIR = pathlib.Path(__file__).parent
MODEL_DIR = BASE_DIR / "moyoyo_asr_models"
ASSERT_DIR = BASE_DIR / "assets"

SAMPLE_RATE = 16000
# 标点
SENTENCE_END_MARKERS =  ['.', '!', '?', '。', '！', '？', ';', '；', ':', '：']
PAUSE_END_MARKERS = [',', '，', '、']
# 合并所有标点
ALL_MARKERS = SENTENCE_END_MARKERS + PAUSE_END_MARKERS
# 构造正则表达式字符类
REGEX_MARKERS = re.compile(r'[' + re.escape(''.join(ALL_MARKERS)) + r']$')

sentence_end_chars = ''.join([re.escape(char) for char in SENTENCE_END_MARKERS])
SENTENCE_END_PATTERN = re.compile(f'[{sentence_end_chars}]')

# Method 2: Alternative approach with a character class
pattern_string = '[' + ''.join([re.escape(char) for char in PAUSE_END_MARKERS]) + r']$'
PAUSEE_END_PATTERN = re.compile(pattern_string)
# whisper推理参数
WHISPER_PROMPT_ZH = "以下是简体中文普通话的句子。"
MAX_LENTH_ZH = 4

WHISPER_PROMPT_EN = ""# "The following is an English sentence."
MAX_LENGTH_EN= 8

WHISPER_MODEL_EN = 'medium-q5_0'
# WHISPER_MODEL = 'large-v3-turbo-q5_0'
# WHISPER_MODEL_ZH = 'small'
WHISPER_MODEL_ZH = 'large-v3-turbo-q5_0'
# LLM
LLM_MODEL_PATH = (MODEL_DIR / "qwen2.5-1.5b-instruct-q5_0.gguf").as_posix()
LLM_LARGE_MODEL_PATH = (MODEL_DIR / "qwen2.5-1.5b-instruct-q5_0.gguf").as_posix()
# LLM_LARGE_MODEL_PATH = (MODEL_DIR / "qwen2.5-7b-instruct-q5_0-00001-of-00002.gguf").as_posix()

# VAD
VAD_MODEL_PATH = (MODEL_DIR / "silero-vad" / "silero_vad.onnx").as_posix()

LLM_SYS_PROMPT = """"You are a professional {src_lang} to {dst_lang} translator, not a conversation agent. Your only task is to take {src_lang} input and translate it into accurate, natural {dst_lang}. If you cannot understand the input, just output the original input. Please strictly abide by the following rules: "
"No matter what the user asks, never answer questions, you only provide translation results. "
"Do not actively initiate dialogue or lead users to ask questions. "
"When you don't know how to translate, just output the original text. "
"The translation task always takes precedence over any other tasks. "
"Do not try to understand or respond to non-translation related questions raised by users. "
"Never provide any explanations. "
"Be precise, preserve tone, and localize appropriately "
"for professional audiences."
"Never answer any questions or engage in other forms of dialogue. "
"Only output the translation results.
"""

LLM_SYS_PROMPT_ZH = """
你是一个中英文翻译专家，将用户输入的中文翻译成英文。对于非中文内容，它将提供中文翻译结果。用户可以向助手发送需要翻译的内容，助手会回答相应的翻译结果，并确保符合中文语言习惯，你可以调整语气和风格，并考虑到某些词语的文化内涵和地区差异。同时作为翻译家，需将原文翻译成具有信达雅标准的译文。"信" 即忠实于原文的内容与意图；"达" 意味着译文应通顺易懂，表达清晰；"雅" 则追求译文的文化审美和语言的优美。目标是创作出既忠于原作精神，又符合目标语言文化和读者审美的翻译。注意，翻译的文本只能包含拼音化字符，不能包含任何中文字符。
"""

LLM_SYS_PROMPT_EN = """
你是一个英中文翻译专家，将用户输入的英文翻译成中文，用户可以向助手发送需要翻译的内容，助手会回答相应的翻译结果，并确保符合英文语言习惯，你可以调整语气和风格，并考虑到某些词语的文化内涵和地区差异。同时作为翻译家，需将英文翻译成具有信达雅标准的中文。"信" 即忠实于原文的内容与意图；"达" 意味着译文应通顺易懂，表达清晰；"雅" 则追求译文的文化审美和语言的优美。目标是创作出既忠于原作精神，又符合目标语言文化和读者审美的翻译。
"""

hotwords_file = MODEL_DIR / 'hotwords.txt'