from lxml import etree
from typing import Any, List, Dict
import numpy as np
import logging
from modules.data import styles_mgr
from modules.speaker import speaker_mgr
import random
logger = logging.getLogger(__name__)
def expand_spk(attrs: dict):
input_spk = attrs.get("spk", "")
if isinstance(input_spk, int):
return
if isinstance(input_spk, str) and input_spk.isdigit():
attrs.update({"spk": int(input_spk)})
return
try:
speaker = speaker_mgr.get_speaker(input_spk)
attrs.update({"spk": speaker})
except Exception as e:
logger.error(f"apply style failed, {e}")
def expand_style(attrs: dict):
if attrs.get("style", "") != "":
try:
params = styles_mgr.find_params_by_name(str(attrs["style"]))
attrs.update(params)
except Exception as e:
logger.error(f"apply style failed, {e}")
def merge_prompt(attrs: dict, elem):
def attr_num(attrs: Dict[str, Any], k: str, min_value: int, max_value: int):
val = elem.get(k, attrs.get(k, ""))
if val == "":
return
if val == "max":
val = max_value
if val == "min":
val = min_value
val = np.clip(int(val), min_value, max_value)
if "prefix" not in attrs or attrs["prefix"] == None:
attrs["prefix"] = ""
attrs["prefix"] += " " + f"[{k}_{val}]"
attr_num(attrs, "oral", 0, 9)
attr_num(attrs, "speed", 0, 9)
attr_num(attrs, "laugh", 0, 2)
attr_num(attrs, "break", 0, 7)
def apply_random_seed(attrs: dict):
seed = attrs.get("seed", "")
if seed == "random" or seed == "rand":
seed = random.randint(0, 2**32 - 1)
attrs["seed"] = seed
logger.info(f"random seed: {seed}")
class NotSupportSSML(Exception):
pass
def parse_ssml(ssml: str) -> List[Dict[str, Any]]:
root = etree.fromstring(ssml)
ssml_version = root.get("version", "NONE")
if ssml_version != "0.1":
raise NotSupportSSML("Unsupported ssml version: {ssml_version}")
segments = []
for voice in root.findall(".//voice"):
voice_attrs = {
"spk": voice.get("spk"),
"style": voice.get("style"),
"seed": voice.get("seed"),
"top_p": voice.get("top_p"),
"top_k": voice.get("top_k"),
"temp": voice.get("temp"),
"prompt1": voice.get("prompt1"),
"prompt2": voice.get("prompt2"),
"prefix": voice.get("prefix"),
"normalize": voice.get("normalize"),
}
voice_attrs = {k: v for k, v in voice_attrs.items() if v is not None}
expand_spk(voice_attrs)
expand_style(voice_attrs)
merge_prompt(voice_attrs, voice)
apply_random_seed(voice_attrs)
voice_segments = []
if voice_attrs.get("temp", "") == "min":
# ref: https://github.com/2noise/ChatTTS/issues/123#issue-2326908144
voice_attrs["temp"] = 0.000000000001
if voice_attrs.get("temp", "") == "max":
voice_attrs["temp"] = 1
# 处理 voice 开头的文本
if voice.text and voice.text.strip():
voice_segments.append(
{"text": voice.text.strip(), "attrs": voice_attrs.copy()}
)
# 处理 voice 内部的文本和 prosody 元素
for node in voice.iterchildren():
if node.tag == "prosody":
prosody_attrs = voice_attrs.copy()
new_attrs = {
"rate": node.get("rate"),
"volume": node.get("volume"),
"pitch": node.get("pitch"),
}
prosody_attrs.update(
{k: v for k, v in new_attrs.items() if v is not None}
)
expand_style(prosody_attrs)
merge_prompt(prosody_attrs, node)
apply_random_seed(voice_attrs)
if node.text and node.text.strip():
voice_segments.append(
{"text": node.text.strip(), "attrs": prosody_attrs}
)
elif node.tag == "break":
time_ms = int(node.get("time", "0").replace("ms", ""))
segment = {"break": time_ms}
voice_segments.append(segment)
if node.tail and node.tail.strip():
voice_segments.append(
{"text": node.tail.strip(), "attrs": voice_attrs.copy()}
)
end_segment = voice_segments[-1]
end_segment["is_end"] = True
segments = segments + voice_segments
logger.info(f"collect len(segments): {len(segments)}")
# logger.info(f"segments: {json.dumps(segments, ensure_ascii=False)}")
return segments
if __name__ == "__main__":
# 示例 SSML 输入
ssml1 = """
电影中梁朝伟扮演的陈永仁的
编号27149
编号27149
编号27149
编号27149
编号27149
电影中梁朝伟扮演的陈永仁的编号27149
"""
ssml2 = """
也可以合成多角色多情感的有声 [uv_break] 书 [uv_break] ，例如：
黛玉冷笑道：
我说呢，亏了绊住，不然，早就飞了来了。
宝玉道：
“只许和你玩，替你解闷。不过偶然到他那里，就说这些闲话。”
“好没意思的话！去不去，关我什么事儿？又没叫你替我解闷儿，还许你不理我呢”
说着，便赌气回房去了。
"""
ssml22 = """
下面是一个 ChatTTS 用于合成多角色多情感的有声书示例
黛玉冷笑道：
我说呢 [uv_break] ，亏了绊住，不然，早就飞起来了。
宝玉道：
“只许和你玩 [uv_break] ，替你解闷。不过偶然到他那里，就说这些闲话。”
“好没意思的话！[uv_break] 去不去，关我什么事儿？ 又没叫你替我解闷儿 [uv_break]，还许你不理我呢”
说着，便赌气回房去了。
"""
ssml3 = """
“你到底在想什么？这已经是第三次了！每次我都告诉你要按时完成任务，可你总是拖延。你知道这对整个团队有多大的影响吗？！”
“你到底在想什么？这已经是第三次了！每次我都告诉你要按时完成任务，可你总是拖延。你知道这对整个团队有多大的影响吗？！”
“你到底在想什么？这已经是第三次了！每次我都告诉你要按时完成任务，可你总是拖延。你知道这对整个团队有多大的影响吗？！”
"""
ssml4 = """
使用 prosody 控制生成文本的语速语调和音量，示例如下
无任何限制将会继承父级voice配置进行生成
设置 rate 大于1表示加速，小于1为减速
设置 pitch 调整音调，设置为6表示提高6个半音
设置 volume 调整音量，设置为2表示提高2个分贝
在 voice 中无prosody包裹的文本即为默认生成状态下的语音
"""
ssml5 = """
使用 break 标签将会简单的
插入一段空白到生成结果中
"""
ssml6 = """
temperature for sampling (may be overridden by style or speaker)
温度值用于采样，这个值有可能被 style 或者 speaker 覆盖
temperature for sampling ，这个值有可能被 style 或者 speaker 覆盖
温度值用于采样，(may be overridden by style or speaker)
"""
segments = parse_ssml(ssml6)
print(segments)
# audio_segments = synthesize_segments(segments)
# combined_audio = combine_audio_segments(audio_segments)
# combined_audio.export("output.wav", format="wav")