#!/usr/bin/env python3 """ 《林夏》TTS 批量生产脚本 用法: conda activate cosyvoice cd /home/xsl/blind python audio/batch_tts.py [--role mom] [--test] 参数: --role ROLE 只生产指定角色（不指定则全部） --test 只合成第一条，用于快速验证环境 --dry-run 只打印任务列表，不实际合成 """ import os, sys, time, argparse, warnings import torch import soundfile as sf # torchaudio nightly requires torchcodec; use soundfile instead warnings.filterwarnings("ignore") COSYVOICE_DIR = "/home/xsl/tools/CosyVoice" PROJECT_DIR = "/home/xsl/blind" REF_DIR = f"{PROJECT_DIR}/audio/00_raw/tts/prompt_ref" OUT_DIR = f"{PROJECT_DIR}/audio/00_raw/tts" sys.path.insert(0, COSYVOICE_DIR) sys.path.insert(0, f"{COSYVOICE_DIR}/third_party/Matcha-TTS") os.chdir(COSYVOICE_DIR) # ────────────────────────────────────────────────────────────── # 角色配置（对应 SOP §4.2） # ref_wav / ref_text 在参考音频备齐前先留空 # ────────────────────────────────────────────────────────────── ROLES = { "mom": { "speed": 0.92, "ref_wav": f"{REF_DIR}/mom_ref.wav", "ref_text": "", # ← 参考音频对应文字，备齐后填入 }, "azhe": { "speed": 1.05, "ref_wav": f"{REF_DIR}/azhe_ref.wav", "ref_text": "", }, "xiaomei": { "speed": 1.10, "ref_wav": f"{REF_DIR}/xiaomei_ref.wav", "ref_text": "", }, "xiaomei_drunk": { # 醉语音用同参考，speed 调慢 "speed": 0.95, "ref_wav": f"{REF_DIR}/xiaomei_ref.wav", "ref_text": "", }, "zhounan": { "speed": 0.95, "ref_wav": f"{REF_DIR}/zhounan_ref.wav", "ref_text": "", }, "hr": { "speed": 1.00, "ref_wav": f"{REF_DIR}/hr_ref.wav", "ref_text": "", }, "anan": { "speed": 1.15, "ref_wav": f"{REF_DIR}/anan_ref.wav", "ref_text": "", }, "dorm_a": { "speed": 1.10, "ref_wav": f"{REF_DIR}/dorm_a_ref.wav", "ref_text": "", }, "dorm_b": { "speed": 1.00, "ref_wav": f"{REF_DIR}/dorm_b_ref.wav", "ref_text": "", }, "dorm_c": { "speed": 0.95, "ref_wav": f"{REF_DIR}/dorm_c_ref.wav", "ref_text": "", }, "delivery": { "speed": 1.20, "ref_wav": f"{REF_DIR}/delivery_ref.wav", "ref_text": "", }, "self_3y": { # 复用 xiaomei，speed +0.05 "speed": 1.15, "ref_wav": f"{REF_DIR}/xiaomei_ref.wav", "ref_text": "", }, # system 用预训练音色，不在这里 } # ────────────────────────────────────────────────────────────── # 台词表（对应林夏.md §3 + §5） # 格式: (输出文件前缀, 角色key, 台词) # 长戏（≥30s）已按 SOP §5.3 拆分为分段 # ────────────────────────────────────────────────────────────── LINES = [ # ── 01 · 21:03 妈妈电话 ── ("lv11_2103_mom_call_01", "mom", "今晚煮了你爱吃的排骨，给你拍张照啊。下周五回来一趟，你高中物理老师的儿子，挺好的。"), # ── 02 · 21:15 小美三连 ── ("lv11_2115_xiaomei_voice_01", "xiaomei", "在吗"), ("lv11_2115_xiaomei_voice_02", "xiaomei", "你被甩了我都听说了"), ("lv11_2115_xiaomei_voice_03", "xiaomei", "明天晚上喝酒，海底捞旁边那家，不许放鸽子"), # ── 03 · 21:20 外卖小哥 ── ("lv11_2120_delivery_call_01", "delivery", "你好你的麻辣烫到了，放门口可以吗？"), # ── 04 · 21:30 阿哲语音 ── ("lv11_2130_azhe_voice_01", "azhe", "嗯……夏夏。我那个AirPods充电盒，好像落你抽屉里了。" "你方便的时候帮我寄一下吗，工位还要用。地址我等下发你。"), # ── 05 · 21:45 HR 王姐 ── ("lv11_2145_hr_voice_01", "hr", "夏夏啊，离职流程你下周三之前办完哈，公积金那边记得跟进。" "辛苦啦，外面有合适的工作我帮你留意。"), # ── 06 · 22:00 安安 ── ("lv11_2200_anan_voice_01", "anan", "夏夏！我在helens！要不要过来！我请客！"), # ── 07 · 22:15 寝室群 ── ("lv11_2215_dorm_a_voice_01", "dorm_a", "哎你们看，我今天升职了！"), ("lv11_2215_dorm_b_voice_01", "dorm_b", "真的啊！恭喜恭喜！太厉害了！"), ("lv11_2215_dorm_c_voice_01", "dorm_c", "厉害！那要请客啊！"), ("lv11_2215_dorm_a_voice_02", "dorm_a", "哈哈哈必须的！夏夏你怎么不说话呀"), ("lv11_2215_dorm_b_voice_02", "dorm_b", "夏夏在忙吧，别打扰她了"), # ── 09 · 22:45 小美醉语音（分 4 段）── ("lv11_2245_xiaomei_voice_01_seg1", "xiaomei_drunk", "夏夏夏夏夏。你睡了吗？我跟你说，我跟你说啊……今天我喝多了……"), ("lv11_2245_xiaomei_voice_01_seg2", "xiaomei_drunk", "你别嫌我烦哈，我跟你说……其实啊……阿哲他，大三那年，他来追过我。" "我没答应他。我那时候就觉得他这人，呵，差点意思。"), ("lv11_2245_xiaomei_voice_01_seg3", "xiaomei_drunk", "你跟他在一起的时候，我没说，因为说了也没意义嘛对吧。" "现在他甩了你，我才敢说。"), ("lv11_2245_xiaomei_voice_01_seg4", "xiaomei_drunk", "夏夏，你听到了吗？你别恨我啊。……喂？算了。"), # ── 10 · 23:00 妈妈文字 ── ("lv11_2300_mom_text_01", "mom", "睡了吗？"), # ── 11 · 23:15 阿哲电话 ── ("lv11_2315_azhe_call_01", "azhe", "我能过去拿吗？就5分钟"), # ── 12 · 23:30 周南第一条 ── ("lv11_2330_zhounan_voice_01", "zhounan", "林夏，是我，周南。你应该不记得我了吧，初三坐你后桌那个"), # ── 13 · 23:45 周南长讲述（4 段）── ("lv11_2345_zhounan_voice_01", "zhounan", "其实我一直想找机会跟你说声谢谢。初三那年你借我笔记，那次考试我才没挂科。"), ("lv11_2345_zhounan_voice_02", "zhounan", "后来高考，我没考好。复读了一年，去了一个二本。然后就回老家了，没什么好说的。"), ("lv11_2345_zhounan_voice_03", "zhounan", "今年第一次来北京出差。来之前我想着，要不要找你。找了很久，才在一个校友群里看到你。" "你现在过得怎么样？"), ("lv11_2345_zhounan_voice_04", "zhounan", "今天我生日。本来就想跟你说一声，对不起这么晚。"), # ── 14 · 00:15 妈妈电话 ── ("lv11_0015_mom_call_01", "mom", "夏夏，你不接电话我有点担心。妈妈知道你工作忙，你早点睡。"), # ── 16 · 02:00 妈妈长语音（分 8 段）── ("lv11_0200_mom_voice_01_seg1", "mom", "夏夏。妈妈知道你今天电话不接。妈妈不催你，就是想跟你说一件事。"), ("lv11_0200_mom_voice_01_seg2", "mom", "你24了对吧。妈妈24那年——你应该没听我说过——"), ("lv11_0200_mom_voice_01_seg3", "mom", "那时候妈妈在广州，给一个香港老板做秘书，1995年。" "那年我也被分手过。也被开除过。一模一样的春天。"), ("lv11_0200_mom_voice_01_seg4", "mom", "那年我哭得特别难看。后来呢？后来就没什么后来了。" "我回了老家，认识了你爸，生了你。"), ("lv11_0200_mom_voice_01_seg5", "mom", "我跟你说这个不是想跟你说什么都会过去——" "我想跟你说，那年夏天我没死，是因为我外婆每天晚上都给我打一个电话。"), ("lv11_0200_mom_voice_01_seg6", "mom", "她不说什么，就问我今天吃饭了没。"), ("lv11_0200_mom_voice_01_seg7", "mom", "我现在打给你，你不接也没事。妈妈给你发个语音，你什么时候想听就听。你不用回我。"), ("lv11_0200_mom_voice_01_seg8", "mom", "你今天接了好多电话吧。妈妈不打扰你了。晚安。"), # ── self-3y 三版（台词待编剧定稿）── # ("lv11_0030_self-3y_voice_01", "self_3y", "（明朗版台词）"), # ("lv11_0030_self-3y_voice_02", "self_3y", "（迷茫版台词）"), # ("lv11_0030_self-3y_voice_03", "self_3y", "（温柔版台词）"), ] def load_model(): from cosyvoice.cli.cosyvoice import CosyVoice2 print("加载 CosyVoice2 模型...") model = CosyVoice2("pretrained_models/CosyVoice2-0.5B") print(f"模型就绪，采样率 {model.sample_rate}Hz") return model def load_ref(role_cfg): from cosyvoice.utils.file_utils import load_wav wav_path = role_cfg["ref_wav"] if not os.path.exists(wav_path): return None, None wav = load_wav(wav_path, 16000) # 裁到 30s 以内 max_samples = 16000 * 30 if wav.shape[1] > max_samples: wav = wav[:, :max_samples] return wav, role_cfg["ref_text"] def synthesize(model, filename, role_key, text): role_cfg = ROLES[role_key] ref_wav, ref_text = load_ref(role_cfg) out_path = f"{OUT_DIR}/{filename}_v1.wav" if os.path.exists(out_path): print(f" [跳过，已存在] {filename}") return True if ref_wav is None: print(f" [跳过，缺参考音频] {role_key}: {role_cfg['ref_wav']}") return False if not ref_text.strip(): print(f" [跳过，ref_text 未填] {role_key}") return False t0 = time.time() chunks = [] for chunk in model.inference_zero_shot( tts_text=text, prompt_text=ref_text, prompt_speech_16k=ref_wav, stream=False, speed=role_cfg["speed"], ): chunks.append(chunk["tts_speech"]) if not chunks: print(f" [失败，无输出] {filename}") return False audio = torch.concat(chunks, dim=1) duration = audio.shape[1] / model.sample_rate sf.write(out_path, audio.squeeze().cpu().numpy(), model.sample_rate) elapsed = time.time() - t0 rtf = elapsed / duration if duration > 0 else 0 print(f" ✓ {filename} {duration:.1f}s 音频 {elapsed:.1f}s 耗时 RTF={rtf:.3f}") return True def main(): parser = argparse.ArgumentParser() parser.add_argument("--role", help="只合成指定角色") parser.add_argument("--test", action="store_true", help="只合成第一条") parser.add_argument("--dry-run", action="store_true", help="只打印任务") args = parser.parse_args() lines = LINES if args.role: lines = [(f, r, t) for f, r, t in lines if r == args.role or r.startswith(args.role)] if args.test: lines = lines[:1] print(f"\n{'='*60}") print(f" 《林夏》TTS 批量合成 共 {len(lines)} 条") print(f"{'='*60}") if args.dry_run: for f, r, t in lines: ref_exists = "✓" if os.path.exists(ROLES.get(r, {}).get("ref_wav", "")) else "✗缺参考" print(f" [{ref_exists}] {f} [{r}] {t[:30]}...") return model = load_model() ok, skip, fail = 0, 0, 0 for i, (filename, role_key, text) in enumerate(lines, 1): print(f"\n[{i}/{len(lines)}] {filename}") result = synthesize(model, filename, role_key, text) if result is True: ok += 1 elif result is None: skip += 1 else: fail += 1 print(f"\n{'='*60}") print(f" 完成：{ok} 条 跳过：{skip} 条 失败：{fail} 条") print(f" 输出目录：{OUT_DIR}") print(f"{'='*60}\n") if __name__ == "__main__": main()