486 lines
24 KiB
Python
486 lines
24 KiB
Python
"""音色自检:用基频(F0)客观判断一段音频实际是男声还是女声。
|
||
|
||
为什么需要这一层:IndexTTS2 这类 zero-shot 克隆模型**只认参考音频**,
|
||
表单里的「性别 / 音色描述」对模型来说是纯元数据。于是出现过很荒唐的一幕——
|
||
用户在平台资产库里把声线性别填成「男声」,因为没上传参考音频,
|
||
生成时回退到模型配置里的那一段默认参考(女声),出来的试听样本是女声,
|
||
界面却全绿、什么都不说。
|
||
|
||
这里做两件事:
|
||
1. ``resolve_reference()``:生成前按性别挑一段**真正能决定音色**的参考音频,
|
||
挑不到就明确报错,绝不静默用默认嗓;
|
||
2. ``check_voice()``:生成后量一次中位基频,与设定性别比对,
|
||
结果写进素材 metadata,界面能直接看到「音色与设定不符」。
|
||
"""
|
||
from __future__ import annotations
|
||
|
||
import os
|
||
import re
|
||
import shutil
|
||
import subprocess
|
||
import tempfile
|
||
from pathlib import Path
|
||
|
||
from sqlalchemy import select
|
||
from sqlalchemy.orm import Session
|
||
|
||
from app.services import storage
|
||
|
||
# 成人基频经验区间:男声 85–155 Hz、女声 165–300 Hz,中间留一段不妄下结论。
|
||
MALE_F0_MAX = 155.0
|
||
FEMALE_F0_MIN = 175.0
|
||
|
||
_SAMPLE_RATE = 8000
|
||
_FRAME = 400 # 50 ms,够容纳 70 Hz 的两个周期
|
||
_HOP = 200 # 25 ms
|
||
_LAG_LO = int(_SAMPLE_RATE / 350) # 最高 350 Hz
|
||
_LAG_HI = int(_SAMPLE_RATE / 70) # 最低 70 Hz
|
||
_MAX_FRAMES = 260 # 纯 Python 逐帧算,长音频均匀抽样即可
|
||
# 折子谐波时「算真谷」的容差:谷深至少要比无周期水平低到全局最深谷的 45% 以内才认。
|
||
_FOLD_TOLERANCE = 0.45
|
||
|
||
GENDER_LABELS = {"male": "男声", "female": "女声", "neutral": "中性声"}
|
||
|
||
#: 声线绑定指向 AI 试听样本时的统一开场白:解释为什么「有声音文件」不等于「有音色」。
|
||
_AI_BOUND_NOTE = "声线现在绑的是 AI 生成的试听样本,它不能反过来当克隆参考(否则错音色会自我强化);"
|
||
|
||
|
||
def _decode(path: Path) -> list[int]:
|
||
"""把任意音频解成 8 kHz 单声道 PCM 整数序列(不引入 numpy 依赖)。"""
|
||
# 必须走项目的 ffmpeg 解析器:本机 ffmpeg 常常不在 PATH 上(配置指定 /
|
||
# 静态构建装在非标准目录)。直接 shutil.which 会静默返回空 → 音色自检
|
||
# 永远「测不出基频」,这个保护就变成了摆设还没人发现。
|
||
from app.services import ffmpeg as ff
|
||
|
||
ffmpeg = ff.FFMPEG if ff.available() else (shutil.which("ffmpeg") or "")
|
||
if not ffmpeg:
|
||
return []
|
||
with tempfile.TemporaryDirectory() as tmp:
|
||
raw = Path(tmp) / "pcm.raw"
|
||
try:
|
||
subprocess.run([ffmpeg, "-y", "-v", "error", "-i", str(path), "-ac", "1",
|
||
"-ar", str(_SAMPLE_RATE), "-f", "s16le", str(raw)],
|
||
check=True, capture_output=True, timeout=120)
|
||
except (subprocess.SubprocessError, OSError):
|
||
return []
|
||
if not raw.exists():
|
||
return []
|
||
import array
|
||
samples = array.array("h")
|
||
samples.frombytes(raw.read_bytes())
|
||
return samples.tolist()
|
||
|
||
|
||
def estimate_f0(path: str | Path) -> float | None:
|
||
"""中位基频(Hz)。取不出可靠音帧时返回 None,而不是猜一个。"""
|
||
produced = Path(path)
|
||
if not produced.exists():
|
||
return None
|
||
samples = _decode(produced)
|
||
if len(samples) < _FRAME * 4:
|
||
return None
|
||
|
||
# 先用整段能量的分位数定音门限,避免把静音/呼吸当音高。
|
||
energies: list[tuple[int, float]] = []
|
||
for start in range(0, len(samples) - _FRAME, _HOP):
|
||
frame = samples[start:start + _FRAME]
|
||
energies.append((start, sum(value * value for value in frame) / _FRAME))
|
||
if not energies:
|
||
return None
|
||
ordered = sorted(energy for _, energy in energies)
|
||
loud = ordered[int(len(ordered) * 0.95)] if len(ordered) > 20 else ordered[-1]
|
||
threshold = max(1500.0, loud * 0.06)
|
||
voiced = [(start, energy) for start, energy in energies if energy >= threshold]
|
||
if len(voiced) < 4:
|
||
return None
|
||
|
||
step = max(1, len(voiced) // _MAX_FRAMES)
|
||
voiced = voiced[::step]
|
||
|
||
pitches: list[float] = []
|
||
for start, _energy in voiced:
|
||
frame = samples[start:start + _FRAME]
|
||
# 平均幅度差函数(AMDF):比自相关省一半算力,对浊音同样稳定。
|
||
# 必须**按参与项数取平均**:lag 越大参与相减的样本越少,直接比总和会让
|
||
# 2 倍 / 3 倍周期的谷“看起来更深”,把 250 Hz 的女声读成 83 Hz 的“男声”。
|
||
amdf: list[float] = []
|
||
for lag in range(_LAG_LO, _LAG_HI):
|
||
total = 0
|
||
terms = 0
|
||
for index in range(0, _FRAME - lag, 2):
|
||
total += abs(frame[index] - frame[index + lag])
|
||
terms += 1
|
||
amdf.append(total / max(terms, 1))
|
||
if len(amdf) < 4:
|
||
continue
|
||
# flat = 这段“没有周期可比”时的平均水平,best = 最深的那个谷。
|
||
# 两者一起用来判音:谷不够深说明这帧是清音/噪声,直接丢。
|
||
flat = sum(amdf) / len(amdf)
|
||
index = min(range(len(amdf)), key=lambda i: amdf[i])
|
||
best = amdf[index]
|
||
if not flat or best >= flat * 0.72:
|
||
continue
|
||
period = _fold_submultiple(amdf, _refine(amdf, index), best, flat)
|
||
pitches.append(_SAMPLE_RATE / period)
|
||
if len(pitches) < 4:
|
||
return None
|
||
pitches.sort()
|
||
return round(pitches[len(pitches) // 2], 1)
|
||
|
||
|
||
def _amdf_at(amdf: list[float], period: float) -> float:
|
||
"""在**非整数周期**上取 AMDF 值(线性插值)。
|
||
|
||
基频几乎不会正好落在整数采样点上(300 Hz @ 8 kHz = 26.67 点),
|
||
只能在相邻两个 lag 之间插,否则永远量不到真正的谷,八度错误就是这么来的。
|
||
"""
|
||
if period < _LAG_LO:
|
||
return float("inf")
|
||
# amdf[0] 对应 lag=_LAG_LO,所以要减掉偏移再取;越界一律贴边而不是抛错。
|
||
last = len(amdf) - 1
|
||
low = min(int(period) - _LAG_LO, last)
|
||
ratio = min(max(period - _LAG_LO - low, 0.0), 1.0) # 落在最后一点时只剩左侧可用
|
||
left = amdf[low]
|
||
right = amdf[low + 1] if low + 1 <= last else left
|
||
return left + (right - left) * ratio
|
||
|
||
|
||
def _fold_submultiple(amdf: list[float], lag: float, best: float, flat: float) -> float:
|
||
"""把「整数倍周期」的谷折回真正的基频周期。
|
||
|
||
周期不是整数个采样点时(300 Hz @ 8 kHz = 26.67 点),3 倍周期恰好落在整数 lag=80 上,
|
||
AMDF 的全局最低谷会出现在 80 而不是 26.67 —— 音高直接低一个八度以上,
|
||
女声被读成男声(实测 300 Hz 被判成 100 Hz),音色自检就白做了。
|
||
|
||
判据:如果最短周期的 1/2、1/3、1/4 处本身也是一个**货真价实的深谷**
|
||
(深度接近全局最深谷、且是局部极小),说明现在这个谷只是它的整数倍。
|
||
用「相对谷深」而不是归一化互相关阈值:真男声在半周期处是**峰**(波形不对称),
|
||
根本过不了这条线,所以不会反过来把男声折成女声。
|
||
"""
|
||
folded = float(lag)
|
||
for _round in range(3): # 折一次后可能还能再折,迭代收敛
|
||
improved = False
|
||
for divisor in (2, 3, 4):
|
||
candidate = folded / divisor
|
||
if candidate < _LAG_LO:
|
||
continue
|
||
depth = _amdf_at(amdf, candidate)
|
||
if depth >= best + _FOLD_TOLERANCE * (flat - best):
|
||
continue # 这里不够深:不是真周期,别折
|
||
# 还得是局部极小,否则只是在两个采样点里挑了个顺眼的
|
||
if any(_amdf_at(amdf, candidate + delta) < depth - 1e-9
|
||
for delta in (-1.0, 1.0)):
|
||
continue
|
||
folded = candidate
|
||
improved = True
|
||
break
|
||
if not improved:
|
||
break
|
||
return folded
|
||
|
||
|
||
def _refine(amdf: list[float], index: int) -> float:
|
||
"""对最深谷做抛物线插值,把周期量到亚采样点精度。
|
||
|
||
8 kHz 下 250 Hz 的周期是 32 点、247.5 Hz 是 32.32 点,两者整数 lag 都是 32;
|
||
不插值就只能报 250.0,边界性别(155/175 Hz 附近)会判得不稳。
|
||
"""
|
||
if index <= 0 or index >= len(amdf) - 1:
|
||
return float(index + _LAG_LO)
|
||
# 抛物线的极值点必须留在相邻三点之间,否则噪声帧会把它甩出整个 lag 区间
|
||
left, middle, right = amdf[index - 1], amdf[index], amdf[index + 1]
|
||
denominator = left - 2 * middle + right
|
||
if denominator <= 1e-9:
|
||
return float(index + _LAG_LO)
|
||
shift = 0.5 * (left - right) / denominator
|
||
shift = max(-1.0, min(1.0, shift))
|
||
return float(index + _LAG_LO + shift)
|
||
|
||
|
||
def gender_of(f0: float | None) -> str:
|
||
if not f0:
|
||
return ""
|
||
if f0 <= MALE_F0_MAX:
|
||
return "male"
|
||
if f0 >= FEMALE_F0_MIN:
|
||
return "female"
|
||
return "neutral" # 落在男女交界区,不硬判
|
||
|
||
|
||
def check_voice(path: str | Path, declared_gender: str = "") -> dict:
|
||
"""量一次音色并和设定比对。返回可直接进 metadata / 前端的结构。"""
|
||
f0 = estimate_f0(path)
|
||
actual = gender_of(f0)
|
||
declared = str(declared_gender or "").strip()
|
||
result: dict = {"f0_hz": f0, "timbre_gender": actual,
|
||
"timbre_gender_label": GENDER_LABELS.get(actual, "未判定"),
|
||
"declared_gender": declared,
|
||
"declared_gender_label": GENDER_LABELS.get(declared, "未指定"),
|
||
"matched": None, "note": ""}
|
||
if f0 is None:
|
||
result["note"] = "未能从试听样本中提取稳定基频,无法自动判定音色性别"
|
||
return result
|
||
if declared in ("", "neutral"):
|
||
result["note"] = f"中位基频 {f0} Hz,听感接近{result['timbre_gender_label'] or '未判定'};设定未限定性别"
|
||
return result
|
||
if actual == "neutral":
|
||
result["note"] = f"中位基频 {f0} Hz,处于男女声交界区,建议试听确认"
|
||
return result
|
||
result["matched"] = actual == declared
|
||
result["note"] = (f"中位基频 {f0} Hz,判定为{result['timbre_gender_label']},"
|
||
f"与设定「{result['declared_gender_label']}」一致" if result["matched"] else
|
||
f"中位基频 {f0} Hz,实际是{result['timbre_gender_label']},"
|
||
f"与设定「{result['declared_gender_label']}」不符:请换一段{result['declared_gender_label']}"
|
||
f"参考音频重新生成")
|
||
return result
|
||
|
||
|
||
def clone_model(spec) -> bool:
|
||
"""这个 TTS 是否靠参考音频决定音色。
|
||
|
||
IndexTTS2(``comfyui_tts``)是 zero-shot 克隆:给它什么参考就是什么嗓子,
|
||
表单里的性别/描述完全不影响。云端音色库反过来——音色由 provider_voice_id 选定,
|
||
本地参考音频可有可无。两种模型不能共用一套「缺参考就报错」的规则。
|
||
"""
|
||
if spec is None:
|
||
return False
|
||
if str(getattr(spec, "adapter", "") or "") == "comfyui_tts":
|
||
return True
|
||
return bool((getattr(spec, "config_json", None) or {}).get("clone_reference"))
|
||
|
||
|
||
def audio_path_or_empty(value: str | None) -> str:
|
||
"""声线记录里的 provider_voice_id 可能是显示名(如「低沉男声」),只有真音频路径才算数。"""
|
||
from app.services.audio import AUDIO_SUFFIXES
|
||
|
||
candidate = str(value or "").strip()
|
||
if not candidate or Path(candidate).suffix.lower() not in AUDIO_SUFFIXES:
|
||
return ""
|
||
absolute = Path(storage.to_absolute(candidate))
|
||
return str(absolute) if absolute.exists() else ""
|
||
|
||
|
||
def _owner_media_rows(db: Session, owner) -> list:
|
||
"""本声线名下的参考/音色素材(平台母版走 PlatformMedia,项目声线走 Asset)。"""
|
||
from app.models import Asset, PlatformMedia, PlatformVoice
|
||
|
||
if isinstance(owner, PlatformVoice):
|
||
return list(db.scalars(
|
||
select(PlatformMedia).where(PlatformMedia.owner_type == "voice",
|
||
PlatformMedia.owner_id == owner.id,
|
||
PlatformMedia.kind.in_(("reference", "voice_sample")))
|
||
).all())
|
||
return list(db.scalars(
|
||
select(Asset).where(Asset.owner_type == "voice", Asset.owner_id == owner.id,
|
||
Asset.kind.in_(("reference", "voice_sample")))
|
||
).all())
|
||
|
||
|
||
#: 模型产出的素材都带这两个记号之一:显示名里的「· AI 音色样本」,或落盘文件名里的 ``-ai`` 段
|
||
#: (``voice-3-voice_sample-ai_89ecda6f.wav``)。用来兜住引入/上架时没带走 generated 标记的老数据。
|
||
#: 刻意不匹配「sample.wav」这种普通上传名:判错的方向虽然保守,但让用户看着自己上传的
|
||
#: 参考音频被告知「没有音色可用」,比错音色自我强化更难解释。
|
||
_AI_DISPLAY_NAME = re.compile(r"(?:·|\|)\s*AI\s|\bAI\s+音色")
|
||
_AI_FILE_NAME = re.compile(r"(?:^|[-_])ai(?:[-_.]|$)", re.IGNORECASE)
|
||
|
||
|
||
def _is_generated(media) -> bool:
|
||
"""这段音频是不是模型生成的,而不是真人录的。
|
||
|
||
先看元数据;老数据(引入/上架时没带走 ``generated``)再按命名兜底。
|
||
判成「模型生成」是保守方向:宁可让用户补一次参考音频,
|
||
也不能让 AI 的嗓门被当成真人参考、继续克隆它自己的输出。
|
||
"""
|
||
meta = getattr(media, "metadata_json", None) or {}
|
||
if meta.get("generated"):
|
||
return True
|
||
if _AI_DISPLAY_NAME.search(str(getattr(media, "name", "") or "")):
|
||
return True
|
||
files = (str(meta.get("original_name") or ""), Path(str(getattr(media, "path", "") or "")).name)
|
||
return any(_AI_FILE_NAME.search(name) for name in files)
|
||
|
||
|
||
def provenance_of(media) -> dict:
|
||
"""引入 / 上架时一起带走的音色出身信息。
|
||
|
||
复制素材时丢掉 ``generated`` 标记,等于把「AI 生成的试听样本」洗白成
|
||
「真人参考音频」:下一次生成就会克隆它自己的输出,错音色从此自我强化、再也翻不出来。
|
||
"""
|
||
if not _is_generated(media):
|
||
return {}
|
||
meta = getattr(media, "metadata_json", None) or {}
|
||
return {"generated": True, "generated_model": meta.get("model") or "",
|
||
"generated_provider": meta.get("provider") or ""}
|
||
|
||
|
||
def _identity(path: str) -> str:
|
||
"""同一个文件的相对/绝对/软链写法归一,比对绑定字段时才认得出是同一个嗓子。"""
|
||
try:
|
||
return os.path.realpath(path)
|
||
except OSError:
|
||
return str(path)
|
||
|
||
|
||
def generated_audio_index(db: Session, owner) -> dict[str, object]:
|
||
"""本声线名下所有「模型生成」的音频文件 → 素材行。"""
|
||
index: dict[str, object] = {}
|
||
for media in _owner_media_rows(db, owner):
|
||
if not _is_generated(media):
|
||
continue
|
||
candidates = [getattr(media, "path", "")]
|
||
candidates += [row.path for row in (getattr(media, "versions", None) or [])]
|
||
for candidate in candidates:
|
||
absolute = audio_path_or_empty(candidate)
|
||
if absolute:
|
||
index[_identity(absolute)] = media
|
||
return index
|
||
|
||
|
||
def own_reference(db: Session, owner, *, detail: dict | None = None) -> str:
|
||
"""这个声线自己能不能决定音色:先看绑定字段,再看它名下的音色素材。
|
||
|
||
**AI 生成的试听样本不能反过来当参考音频。** 生成样本的音色本来就是这次要定的结论,
|
||
再拿它当输入就成了自我强化的循环:一旦某次因为缺参考出了个女声,
|
||
下一次生成会把这个女声样本当「本声线自己的嗓子」继续克隆,怎么重试都是女声。
|
||
用户上传的样本不在其列——那本来就是真人录音,可以正当当参考。
|
||
|
||
``detail`` 传字典时会记下「声线当前绑着的其实就是 AI 样本」,
|
||
界面才能解释「我明明有声音文件,为什么说我没有音色」。
|
||
"""
|
||
generated = generated_audio_index(db, owner)
|
||
bound = str(getattr(owner, "provider_voice_id", "") or "").strip()
|
||
direct = audio_path_or_empty(bound)
|
||
# 「绑的是不是 AI 样本」跟文件此刻还在不在无关:样本被删了也要照样说清楚,
|
||
# 否则界面只显示「没有任何音色」,用户看不懂自己挂上去的那段声音去哪了。
|
||
bound_is_ai = bool(bound) and (_identity(bound) in generated
|
||
or bool(_AI_FILE_NAME.search(Path(bound).name)))
|
||
if bound_is_ai:
|
||
if detail is not None:
|
||
detail["bound_generated"] = True
|
||
elif direct:
|
||
return direct
|
||
rows = sorted(_owner_media_rows(db, owner), key=lambda row: row.id, reverse=True)
|
||
for media in rows: # 先认真正的参考音频
|
||
if _is_generated(media):
|
||
continue
|
||
path = audio_path_or_empty(getattr(media, "path", ""))
|
||
if path:
|
||
return path
|
||
return ""
|
||
|
||
|
||
def _imported_master(db: Session, owner):
|
||
"""项目声线是从平台母版引入的时,返回那条母版;否则 None。
|
||
|
||
引入过程可能只带走了设定、没带走参考音频。这种情况下最像这个角色的是**它自己的母版**,
|
||
而不是随便借一个同性别的路人嗓子。
|
||
"""
|
||
from app.models import PlatformVoice as _PV
|
||
|
||
source_id = getattr(owner, "platform_source_id", None)
|
||
if (source_id and getattr(owner, "platform_source_type", "") == "voice"
|
||
and not isinstance(owner, _PV)):
|
||
return db.get(_PV, source_id)
|
||
return None
|
||
|
||
|
||
def _same_gender_voices(db: Session, owner, gender: str) -> list:
|
||
"""库里同性别、且真的带着参考音频的其他声线:先项目内,再平台母版。"""
|
||
from app.models import PlatformVoice as _PV
|
||
from app.models import VoiceAsset as _VA
|
||
|
||
peers: list = []
|
||
project_id = getattr(owner, "project_id", None)
|
||
if project_id and not isinstance(owner, _PV):
|
||
peers += list(db.scalars(
|
||
select(_VA).where(_VA.project_id == project_id, _VA.id != owner.id,
|
||
_VA.gender == gender)
|
||
).all())
|
||
peers += [row for row in db.scalars(
|
||
select(_PV).where(_PV.gender == gender, _PV.is_active.is_(True))
|
||
).all() if not isinstance(owner, _PV) or row.id != owner.id]
|
||
return sorted(peers, key=lambda row: (not getattr(row, "is_default", False), -row.id))
|
||
|
||
|
||
def resolve_reference(db: Session, owner, *, explicit: str = "", clone: bool = True) -> dict:
|
||
"""挑一段真正能决定音色的参考音频,并说明它是谁。
|
||
|
||
优先级:人工指定 → 本声线自己的参考 → 引入自的平台母版 → **同性别的其他声线** → 模型默认。
|
||
关键在倒数第二级:设定了性别却没有任何同性别参考时,必须报错,
|
||
不能让「男声」悄悄用上默认女声底模(这正是平台上「男性声音生成成女声」的成因)。
|
||
|
||
``clone=False`` 表示这个 TTS 走云端音色库(provider_voice_id 是音色 ID,不是参考音频),
|
||
此时没有本地参考音频不算问题,交回供应商音色 ID 处理。
|
||
"""
|
||
chosen = audio_path_or_empty(explicit)
|
||
if not clone:
|
||
# 云端音色库靠 provider_voice_id 选音色,本地参考**不能**顶替它:
|
||
# 把 WAV 路径当音色 ID 发过去,macos_say 之类会直接拿到一个不存在的嗓门。
|
||
# 但用户手动挂上来的参考仍然记进 attached,供素材元数据回溯用了哪段音频。
|
||
return {"path": "", "attached": chosen, "source": "provider_voice",
|
||
"note": ("该模型用供应商音色 ID 选音色,本地参考音频只作记录、不决定音色"
|
||
if chosen else "该模型用供应商音色 ID 选音色,不依赖本地参考音频")}
|
||
|
||
if chosen:
|
||
return {"path": chosen, "attached": chosen, "source": "manual",
|
||
"note": "使用你在生成表单里选定的参考音频"}
|
||
|
||
detail: dict = {}
|
||
own = own_reference(db, owner, detail=detail)
|
||
if own:
|
||
return {"path": own, "source": "self", "bound_generated": False,
|
||
"note": "使用本声线已绑定的参考音频"}
|
||
bound_generated = bool(detail.get("bound_generated"))
|
||
|
||
master = _imported_master(db, owner)
|
||
if master is not None:
|
||
master_path = own_reference(db, master)
|
||
if master_path:
|
||
return {"path": master_path, "source": "platform_master",
|
||
"voice_name": master.name, "bound_generated": bound_generated,
|
||
"note": (_AI_BOUND_NOTE + f"本次沿用它的平台母版「{master.name}」的音色"
|
||
if bound_generated else
|
||
"本声线还没有带自己的参考音频,本次沿用它的平台母版"
|
||
f"「{master.name}」的音色")}
|
||
|
||
gender = str(getattr(owner, "gender", "") or "").strip()
|
||
if gender in GENDER_LABELS and gender != "neutral":
|
||
for sibling in _same_gender_voices(db, owner, gender):
|
||
path = own_reference(db, sibling)
|
||
if path:
|
||
return {"path": path, "source": "gender_sibling", "voice_name": sibling.name,
|
||
"bound_generated": bound_generated,
|
||
"note": (_AI_BOUND_NOTE +
|
||
f"本次以同性别声线「{sibling.name}」的音色为底生成;"
|
||
f"想固定成自己的嗓子,请上传一段{GENDER_LABELS[gender]}真人参考音频"
|
||
if bound_generated else
|
||
f"本声线还没有参考音频,本次以同性别声线「{sibling.name}」的音色为底生成;"
|
||
f"想固定成自己的嗓子,请上传一段{GENDER_LABELS[gender]}参考音频")}
|
||
return {"path": "", "source": "none", "bound_generated": bound_generated,
|
||
"note": (_AI_BOUND_NOTE + f"库里也没有任何{GENDER_LABELS[gender]}参考音频可用:"
|
||
f"请上传一段{GENDER_LABELS[gender]}真人参考音频再生成"
|
||
if bound_generated else
|
||
f"库里没有任何{GENDER_LABELS[gender]}参考音频可用:"
|
||
f"请上传一段{GENDER_LABELS[gender]}参考音频再生成,"
|
||
f"否则会拿到与设定不符的音色")}
|
||
return {"path": "", "source": "model_default", "bound_generated": bound_generated,
|
||
"note": (_AI_BOUND_NOTE + "本次使用模型配置的默认参考音频(音色不受表单控制)"
|
||
if bound_generated else
|
||
"声线未限定性别,本次使用模型配置的默认参考音频(音色不受表单控制)")}
|
||
|
||
|
||
def reference_status(db: Session, owner, *, clone: bool = True) -> dict:
|
||
"""给界面用的预告:现在点「生成」会拿到谁的嗓子,值不值得先传参考音频。"""
|
||
resolved = resolve_reference(db, owner, clone=clone)
|
||
gender = str(getattr(owner, "gender", "") or "").strip()
|
||
return {
|
||
"gender": gender,
|
||
"gender_label": GENDER_LABELS.get(gender, "未设定"),
|
||
"source": resolved["source"],
|
||
"note": resolved["note"],
|
||
"borrowed": resolved["source"] in ("gender_sibling", "platform_master"),
|
||
# 用户视角的「我明明传过声音」:绑定的是 AI 样本时界面必须说清楚为什么不算
|
||
"bound_generated": bool(resolved.get("bound_generated")),
|
||
"ready": (bool(resolved["path"]) or resolved["source"] in ("model_default", "provider_voice")),
|
||
}
|