Spaces:
Sleeping
Sleeping
| #!/usr/bin/env python3 | |
| # -*- coding: utf-8 -*- | |
| """原创五声音阶配乐、低量音效、水声及中文 TTS 混音（48kHz 双声道）。""" | |
| import json | |
| import math | |
| import subprocess | |
| import wave | |
| from pathlib import Path | |
| import numpy as np | |
| ROOT = Path(__file__).resolve().parents[1] | |
| STORY = json.loads((ROOT / "story.json").read_text(encoding="utf-8")) | |
| TIMING = json.loads((ROOT / "timing.json").read_text(encoding="utf-8")) | |
| SR = 48000 | |
| N = round(TIMING["duration"] * SR) | |
| rng = np.random.default_rng(20261010) # 固定种子，输出可重复。 | |
| voice = np.zeros((N, 2), dtype=np.float32) | |
| score = np.zeros((N, 2), dtype=np.float32) | |
| fx = np.zeros((N, 2), dtype=np.float32) | |
| def add(dest: np.ndarray, samples: np.ndarray, start_s: float, gain=1.0, pan=0.0) -> None: | |
| start = int(round(start_s * SR)) | |
| if start < 0 or start >= N: | |
| return | |
| length = min(len(samples), N - start) | |
| samples = np.asarray(samples[:length], dtype=np.float32) * gain | |
| dest[start:start + length, 0] += samples * math.sqrt((1.0 - pan) / 2) | |
| dest[start:start + length, 1] += samples * math.sqrt((1.0 + pan) / 2) | |
| def pluck(freq: float, seconds=2.0) -> np.ndarray: | |
| t = np.arange(round(seconds * SR), dtype=np.float32) / SR | |
| ph = 2 * np.pi * freq * t | |
| return ((np.sin(ph) + .27 * np.sin(2 * ph) + .08 * np.sin(3 * ph)) * | |
| np.exp(-2.0 * t) * np.clip(t / .008, 0, 1)).astype(np.float32) | |
| def flute(freq: float, seconds=2.3) -> np.ndarray: | |
| t = np.arange(round(seconds * SR), dtype=np.float32) / SR | |
| ph = 2 * np.pi * freq * t + .04 * np.sin(2 * np.pi * 5.0 * t) | |
| fade = np.clip(t / .14, 0, 1) * np.clip((seconds - t) / .65, 0, 1) | |
| return ((np.sin(ph) + .17 * np.sin(2 * ph)) * fade).astype(np.float32) | |
| def add_noise(at: float, seconds: float, gain: float, freq_smooth=18) -> None: | |
| n = round(seconds * SR) | |
| z = rng.normal(0, 1, n).astype(np.float32) | |
| if freq_smooth > 1: | |
| z = np.convolve(z, np.ones(freq_smooth, dtype=np.float32) / freq_smooth, "same") | |
| t = np.arange(n, dtype=np.float32) / SR | |
| env = np.exp(-t * 4.0) * np.clip(t / .012, 0, 1) | |
| add(fx, z * env, at, gain=gain, pan=.13) | |
| melody = [293.66, 329.63, 392.0, 440.0, 392.0, 329.63, 293.66, 261.63] | |
| for i, start in enumerate(np.arange(0.3, TIMING["duration"] - 1.0, 1.38)): | |
| add(score, pluck(melody[i % len(melody)], 2.4), float(start), gain=.049, | |
| pan=((-1) ** i) * .18) | |
| if i % 4 == 0: | |
| add(score, pluck(melody[i % len(melody)] / 2, 3.3), float(start), gain=.034, pan=-.23) | |
| if i % 5 == 1: | |
| add(score, flute(melody[(i + 3) % len(melody)], 2.5), float(start) + .15, | |
| gain=.017, pan=.24) | |
| for sc, seg in zip(STORY["scenes"], TIMING["scenes"]): | |
| event = sc.get("sfx", "") | |
| at = float(seg["start"]) + min(2.4, max(0.6, seg["duration"] * .39)) | |
| if event == "splash": | |
| add_noise(at, 1.7, gain=.46, freq_smooth=15) | |
| # 水面扩散，低频音效含能量小，避免喧宾夺主。 | |
| t = np.arange(round(.6 * SR), dtype=np.float32) / SR | |
| ripple = (np.sin(2 * np.pi * (48 - 20 * t) * t) * np.exp(-8 * t)).astype(np.float32) | |
| add(fx, ripple, at, gain=.045) | |
| elif event in ("whoosh", "ripple"): | |
| add_noise(at, .85, gain=.18, freq_smooth=55 if event == "ripple" else 27) | |
| elif event == "sparkle": | |
| add(fx, flute(780, 1.7), at, gain=.045, pan=.23) | |
| # 将每个 TTS 语音转成统一的单声道浮点格式，再按真实开始时间叠加。 | |
| for scene in TIMING["scenes"]: | |
| # 每幕可有任意数量的角色对白：使用逐段、实测的 TTS 时间轴混音。 | |
| # 兼容旧版单句场景；不能同时混两套，否则会导致旁白被重复播放。 | |
| lines = scene.get("lines") or [scene] | |
| for seg in lines: | |
| source = ROOT / seg["voice_file"] | |
| raw = subprocess.check_output([ | |
| "ffmpeg", "-hide_banner", "-loglevel", "error", "-i", str(source), | |
| "-ac", "1", "-ar", str(SR), "-f", "f32le", "-" | |
| ]) | |
| audio = np.frombuffer(raw, dtype="<f4").copy() | |
| if not len(audio): | |
| raise ValueError(f"空语音: {source} / 角色：{seg.get('role', 'narrator')}") | |
| amp = float(np.max(np.abs(audio))) | |
| gain = min(1.7, .63 / max(amp, .0001)) | |
| add(voice, audio, float(seg["voice_start"]), gain=gain, pan=0) | |
| # 配乐整体偏低，语音活跃时额外自动压低背景；在片头片尾淡入淡出。 | |
| voice_rms = np.sqrt(np.mean(voice.astype(np.float64) ** 2, axis=1)) | |
| active = voice_rms > .009 | |
| # 近似语音包络，与语音交叉处更平顺。 | |
| gate = active.astype(np.float32) | |
| width = 7200 | |
| # 使用前缀和 O(N) 平滑，避免百万样本 × 7200 的昂贵直接卷积。 | |
| cumulative = np.concatenate((np.zeros(1, dtype=np.float64), np.cumsum(gate, dtype=np.float64))) | |
| idx = np.arange(N) | |
| lo = np.maximum(idx - width // 2, 0) | |
| hi = np.minimum(idx + width // 2, N) | |
| smooth_gate = ((cumulative[hi] - cumulative[lo]) / np.maximum(hi - lo, 1)).astype(np.float32) | |
| score *= (.84 - .26 * np.clip(smooth_gate, 0, 1))[:, None] | |
| times = np.arange(N, dtype=np.float32) / SR | |
| fade = np.minimum(1.0, times / .72) * np.minimum(1.0, (TIMING["duration"] - times) / 1.6) | |
| mix = (voice + score + fx) * fade[:, None] | |
| max_abs = float(np.max(np.abs(mix))) | |
| if max_abs > .91: | |
| mix *= .91 / max_abs | |
| out = ROOT / "assets/film-mix.wav" | |
| out.parent.mkdir(exist_ok=True) | |
| with wave.open(str(out), "wb") as f: | |
| f.setnchannels(2) | |
| f.setsampwidth(2) | |
| f.setframerate(SR) | |
| f.writeframes((mix * 32767).astype("<i2").tobytes()) | |
| print(json.dumps({"audio": str(out), "seconds": TIMING["duration"], | |
| "mock_voice": TIMING.get("mock", False), "peak_before_limiter": round(max_abs, 4)}, | |
| ensure_ascii=False, indent=2)) | |