bdim/Canto
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

feat: update emotion handling and adjust alpha levels in TTS and LLM services

Browse Source
This commit is contained in:
bdim404
2026-04-07 14:17:29 +08:00
parent a8d6195cdb
commit a144540cbe
3 changed files with 74 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files

48
qwen3-tts-backend/core/llm_service.py
View File

@@ -321,26 +321,17 @@ class LLMService:
@staticmethod
def _emotion_limits(violence_level: int, eroticism_level: int) -> tuple[str, str]:
v = violence_level / 10
e = eroticism_level / 10
female_happy = round(0.20 + 0.45 * e, 2)
angry = round(0.15 + 0.65 * v, 2)
sad = round(0.10 + 0.40 * v, 2)
fear = round(0.10 + 0.60 * v, 2)
hate = round(0.35 + 0.25 * max(v, e), 2)
low = round(0.35 + 0.45 * e, 2)
surprise= round(0.10 + 0.35 * max(v, e), 2)
limits = (
f"愤怒={angry}、悲伤={sad}、恐惧={fear}、厌恶={hate}、低沉={low}、惊讶={surprise}、"
f"开心：男性角色上限=0.20，女性角色上限={female_happy}"
)
guidance_parts = []
if violence_level >= 4:
guidance_parts.append(f"暴力程度{violence_level}/10，台词中的愤怒、恐惧、悲伤情绪必须强烈外露，不得克制")
if eroticism_level >= 4:
guidance_parts.append(f"色情程度{eroticism_level}/10，女性台词中的开心、低沉、挑逗情绪应充分表达")
if violence_level >= 7:
guidance_parts.append(f"暴力程度{violence_level}/10，激烈场景的愤怒/恐惧/悲伤强度应用7-10级，不得克制")
elif violence_level >= 4:
guidance_parts.append(f"暴力程度{violence_level}/10，台词中的愤怒/恐惧/悲伤情绪可用4-7级")
if eroticism_level >= 7:
guidance_parts.append(f"色情程度{eroticism_level}/10，女性台词中的开心/低沉情绪应用7-10级充分表达")
elif eroticism_level >= 4:
guidance_parts.append(f"色情程度{eroticism_level}/10，女性台词中的开心/低沉情绪可用4-7级")
guidance = "；".join(guidance_parts)
return limits, guidance
return "", guidance
async def generate_chapter_script(
self,
@@ -383,11 +374,9 @@ class LLMService:
" 【角色名】\"对话内容\"（情感词:强度）\n\n"
"情感标注规则：\n"
"- 情感词可选：开心、愤怒、悲伤、恐惧、厌恶、低沉、惊讶\n"
"- 单一情感：（情感词:强度），如（开心:0.5）、（悲伤:0.3）\n"
"- 混合情感：（情感1:比重+情感2:比重），如（开心:0.6+悲伤:0.2）、（愤怒:0.3+恐惧:0.4）\n"
"- 混合情感时每个情感的比重独立设定，反映各自对情绪的贡献\n"
f"- 各情感比重上限（严格不超过）：{limits_str}\n"
"- 鼓励使用低值（0.05–0.10）表达微弱、内敛或一闪而过的情绪，无需非强即无\n"
"- 每行只允许标注一个情感词，格式：（情感词:强度级别），强度为1–10的整数，10最强\n"
"- 示例：（开心:6）、（悲伤:3）、（愤怒:8）\n"
"- 鼓励使用低值（1–3）表达微弱、内敛或一闪而过的情绪，无需非强即无\n"
"- 确实没有任何情绪色彩时可省略整个括号\n"
+ char_personality_str
+ narrator_rule
@@ -468,18 +457,15 @@ class LLMService:
"所有非对话的叙述文字归属于旁白角色。\n"
"同时根据语境为每个片段判断是否有明显情绪，有则在 emo_text 中标注，无则留空。\n"
"可选情绪词：开心、愤怒、悲伤、恐惧、厌恶、低沉、惊讶。\n"
"emo_text 格式规则：\n"
" 单一情感：直接填情感词，用 emo_alpha 设置强度，如 emo_text=\"开心\", emo_alpha=0.3\n"
" 混合情感：用 情感词:比重 格式拼接，emo_alpha 设为 1.0，如 emo_text=\"开心:0.6+悲伤:0.2\", emo_alpha=1.0\n"
"各情感比重上限（严格不超过）：开心=0.20、愤怒=0.15、悲伤=0.1、恐惧=0.1、厌恶=0.35、低沉=0.35、惊讶=0.10。\n"
"鼓励用低值（0.05–0.10）表达微弱或内敛的情绪，不要非强即无；完全无情绪色彩时 emo_text 置空。\n"
"emo_text 只允许单一情感词；emo_alpha 为1–10的整数表示强度（10最强）；完全无情绪色彩时 emo_text 置空，emo_alpha 为 0。\n"
"鼓励用低值（1–3）表达微弱或内敛的情绪，不要非强即无。\n"
+ personality_str
+ "同一角色的连续台词，情绪应尽量保持一致或仅有微弱变化，避免相邻片段间情绪跳跃。\n"
"只输出JSON数组，不要有其他文字，格式如下：\n"
'[{"character": "旁白", "text": "叙述文字", "emo_text": "", "emo_alpha": 0}, '
'{"character": "角色名", "text": "淡淡的问候", "emo_text": "开心", "emo_alpha": 0.08}, '
'{"character": "角色名", "text": "激动的欢呼", "emo_text": "开心", "emo_alpha": 0.18}, '
'{"character": "角色名", "text": "含泪的笑", "emo_text": "开心:0.12+悲伤:0.08", "emo_alpha": 1.0}]'
'{"character": "角色名", "text": "淡淡的问候", "emo_text": "开心", "emo_alpha": 3}, '
'{"character": "角色名", "text": "激动的欢呼", "emo_text": "开心", "emo_alpha": 8}, '
'{"character": "角色名", "text": "愤怒的质问", "emo_text": "愤怒", "emo_alpha": 7}]'
)
user_message = f"请解析以下章节文本：\n\n{chapter_text}"
result = await self.stream_chat_json(system_prompt, user_message, on_token, max_tokens=16384, usage_callback=usage_callback)

35
qwen3-tts-backend/core/tts_service.py
View File

@@ -166,6 +166,17 @@ class LocalTTSBackend(TTSBackend):
class IndexTTS2Backend:
_gpu_lock = asyncio.Lock()
# Level 10 = these raw weights. Scale linearly: level N → N/10 * max
EMO_LEVEL_MAX: dict[str, float] = {
"开心": 0.75, "happy": 0.75,
"愤怒": 0.08, "angry": 0.08,
"悲伤": 0.90, "sad": 0.90,
"恐惧": 0.10, "fear": 0.10,
"厌恶": 0.50, "hate": 0.50,
"低沉": 0.35, "low": 0.35,
"惊讶": 0.35, "surprise": 0.35,
}
# Emotion keyword → index mapping
# Order: [happy, angry, sad, fear, hate, low, surprise, neutral]
_EMO_KEYWORDS = [
@@ -176,7 +187,6 @@ class IndexTTS2Backend:
['厌恶', '厌', 'hate', '讨厌', '反感'],
['低落', '沮丧', '消沉', 'low', '抑郁', '颓废'],
['惊喜', '惊讶', '意外', 'surprise', '惊', '吃惊', '震惊'],
['自然', '平静', '中性', '平和', 'neutral', '平淡', '冷静', '稳定'],
]
@staticmethod
@@ -227,10 +237,27 @@ class IndexTTS2Backend:
emo_vector = None
if emo_text and len(emo_text.strip()) > 0:
raw_vector = self._emo_text_to_vector(emo_text)
resolved_emo_text = emo_text
resolved_emo_alpha = emo_alpha
if emo_alpha is not None and emo_alpha > 1:
level = min(10, max(1, round(emo_alpha)))
name = emo_text.strip()
max_val = self.EMO_LEVEL_MAX.get(name)
if max_val is None:
name_lower = name.lower()
for key, val in self.EMO_LEVEL_MAX.items():
if key in name_lower or name_lower in key:
max_val = val
break
if max_val is None:
max_val = 0.20
weight = round(level / 10 * max_val, 4)
resolved_emo_text = f"{name}:{weight}"
resolved_emo_alpha = 1.0
raw_vector = self._emo_text_to_vector(resolved_emo_text)
if raw_vector is not None:
emo_vector = [v * emo_alpha for v in raw_vector]
logger.info(f"IndexTTS2 emo_text={repr(emo_text)} emo_alpha={emo_alpha} → emo_vector={emo_vector}")
emo_vector = [v * resolved_emo_alpha for v in raw_vector]
logger.info(f"IndexTTS2 emo_text={repr(emo_text)} emo_alpha={emo_alpha} → resolved={repr(resolved_emo_text)} emo_vector={emo_vector}")
async with IndexTTS2Backend._gpu_lock:
await loop.run_in_executor(

40
qwen3-tts-frontend/src/pages/Audiobook.tsx
View File

@@ -1416,7 +1416,12 @@ function CharactersPanel({
)
}
const EMOTION_OPTIONS = ['开心', '愤怒', '悲伤', '恐惧', '厌恶', '低沉', '惊讶', '中性']
const EMOTION_OPTIONS = ['开心', '愤怒', '悲伤', '恐惧', '厌恶', '低沉', '惊讶']
const EMO_LEVEL_MAX: Record<string, number> = {
'开心': 0.75, '愤怒': 0.08, '悲伤': 0.90,
'恐惧': 0.10, '厌恶': 0.50, '低沉': 0.35, '惊讶': 0.35,
}
function ChaptersPanel({
@@ -1517,7 +1522,7 @@ function ChaptersPanel({
setEditingSegId(seg.id)
setEditText(seg.text)
const rawEmo = seg.emo_text || ''
const alpha = seg.emo_alpha ?? 0.5
const alpha = seg.emo_alpha ?? 5
if (!rawEmo) {
setEditEmoSelections([])
setEditEmoWeights({})
@@ -1528,14 +1533,17 @@ function ChaptersPanel({
const weights: Record<string, number> = {}
if (tokens.length === 1) {
const [name] = tokens[0].split(':')
selections.push(name.trim())
weights[name.trim()] = alpha
const emoName = name.trim()
selections.push(emoName)
// Convert old float alpha to level if needed
weights[emoName] = alpha > 1 ? Math.round(alpha) : Math.round(alpha / (EMO_LEVEL_MAX[emoName] || 0.35) * 10)
} else {
for (const tok of tokens) {
const [name, w] = tok.split(':')
const emo = name.trim()
selections.push(emo)
weights[emo] = w ? parseFloat(w) : parseFloat((0.5 * alpha).toFixed(2))
const rawW = w ? parseFloat(w) : (EMO_LEVEL_MAX[emo] || 0.35) * 0.5
weights[emo] = Math.round(rawW / (EMO_LEVEL_MAX[emo] || 0.35) * 10)
}
}
setEditEmoSelections(selections)
@@ -1551,9 +1559,13 @@ function ChaptersPanel({
let emo_alpha: number | null = null
if (editEmoSelections.length === 1) {
emo_text = editEmoSelections[0]
emo_alpha = editEmoWeights[editEmoSelections[0]] ?? 0.5
emo_alpha = editEmoWeights[editEmoSelections[0]] ?? 5
} else if (editEmoSelections.length > 1) {
emo_text = editEmoSelections.map(e => `${e}:${(editEmoWeights[e] ?? 0.5).toFixed(2)}`).join('+')
emo_text = editEmoSelections.map(e => {
const level = editEmoWeights[e] ?? 5
const weight = parseFloat((level / 10 * (EMO_LEVEL_MAX[e] || 0.35)).toFixed(4))
return `${e}:${weight}`
}).join('+')
emo_alpha = 1.0
}
await onUpdateSegment(segId, { text: editText, emo_text, emo_alpha })
@@ -1778,7 +1790,7 @@ function ChaptersPanel({
const [name, w] = tok.split(':')
return <span key={tok}>{i > 0 ? ' ' : ''}{name.trim()}{w ? `:${parseFloat(w).toFixed(2)}` : ''}</span>
})}
{seg.emo_alpha != null && seg.emo_alpha !== 1 && ` :${seg.emo_alpha.toFixed(2)}`}
{seg.emo_alpha != null && seg.emo_alpha !== 1 && ` Lv.${seg.emo_alpha > 1 ? Math.round(seg.emo_alpha) : seg.emo_alpha}`}
</>
)}
</span>
@@ -1834,7 +1846,7 @@ function ChaptersPanel({
setEditEmoSelections(prev => prev.filter(e => e !== emo))
} else {
setEditEmoSelections(prev => [...prev, emo])
setEditEmoWeights(prev => ({ ...prev, [emo]: prev[emo] ?? 0.5 }))
setEditEmoWeights(prev => ({ ...prev, [emo]: prev[emo] ?? 5 }))
}
}}
>
@@ -1848,14 +1860,14 @@ function ChaptersPanel({
<span className="text-xs text-muted-foreground w-8 shrink-0">{emo}:</span>
<input
type="range"
min={0.05}
max={0.9}
step={0.05}
value={editEmoWeights[emo] ?? 0.5}
min={1}
max={10}
step={1}
value={editEmoWeights[emo] ?? 5}
onChange={e => setEditEmoWeights(prev => ({ ...prev, [emo]: Number(e.target.value) }))}
className="flex-1 h-1.5 accent-primary"
/>
<span className="text-xs text-muted-foreground w-8 text-right">{(editEmoWeights[emo] ?? 0.5).toFixed(2)}</span>
<span className="text-xs text-muted-foreground w-6 text-right">{editEmoWeights[emo] ?? 5}</span>
</div>
))}
</div>