移除 GPT-SoVITS 和 Fish Speech，整理 audio_1 脚本与文档

- 删除 03_gen_tts_gptsovits.py、00_transcribe_refs.py、ref_texts.json
- 删除 02_gen_tts_fishspeech.py 及 tts/v2_fishspeech/ 生成结果
- 保留 VoxCPM2 / ACE-Step / Stable Audio 三条链路
- 更新执行文档和技术报告，清除两款 TTS 的所有引用

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

audio_1/scripts/01_gen_tts_voxcpm2.py

@@ -0,0 +1,150 @@
+#!/usr/bin/env python3
+"""
+步骤 1：TTS v1 — VoxCPM2 批量生成
+输出目录：audio_1/tts/v1_voxcpm2/
+
+使用方法：
+  # 确认 VoxCPM2 服务已启动
+  python audio_1/scripts/01_gen_tts_voxcpm2.py
+  python audio_1/scripts/01_gen_tts_voxcpm2.py --dry-run
+  python audio_1/scripts/01_gen_tts_voxcpm2.py --role linxia
+"""
+import os, sys, glob, base64, json, tempfile, time, subprocess, argparse
+import requests
+
+PROJECT_DIR = "/home/xsl/blind"
+VOICE_DIR   = f"{PROJECT_DIR}/voice"
+OUT_DIR     = f"{PROJECT_DIR}/audio_1/tts/v1_voxcpm2"
+HOST = "127.0.0.1"
+PORT = "8000"
+BASE_URL = f"http://{HOST}:{PORT}"
+
+os.makedirs(OUT_DIR, exist_ok=True)
+
+sys.path.insert(0, f"{PROJECT_DIR}/audio")
+# 复用原始脚本的 ROLES 和 LINES 定义
+from batch_tts_voxcpm import ROLES, LINES
+
+# ──────────────────────────────────────────────────────────────
+def find_ref_audio(role_key):
+    role_dir = os.path.join(VOICE_DIR, ROLES[role_key]["dir"])
+    for ext in ("*.wav", "*.mp3", "*.flac"):
+        files = sorted(glob.glob(os.path.join(role_dir, ext)))
+        if files:
+            return files[0]
+    return None
+
+def wav_to_base64(path):
+    with open(path, "rb") as f:
+        return base64.b64encode(f.read()).decode()
+
+def convert_to_wav(src, dst):
+    subprocess.run(["ffmpeg", "-y", "-i", src, "-ar", "16000", "-ac", "1",
+                    "-acodec", "pcm_s16le", dst],
+                   check=True, capture_output=True)
+
+def apply_speed(src, speed, dst):
+    if abs(speed - 1.0) < 0.01:
+        import shutil; shutil.copy2(src, dst); return
+    subprocess.run(["ffmpeg", "-y", "-i", src, "-filter:a", f"atempo={speed}",
+                    "-acodec", "pcm_s16le", dst],
+                   check=True, capture_output=True)
+
+def wav_to_mp3(src, dst):
+    subprocess.run(["ffmpeg", "-y", "-i", src, "-codec:a", "libmp3lame",
+                    "-b:a", "128k", "-ar", "24000", "-ac", "1", dst],
+                   check=True, capture_output=True)
+
+def check_server():
+    try:
+        return requests.get(f"{BASE_URL}/health", timeout=5).status_code == 200
+    except: return False
+
+def register_voice(wav_path):
+    r = requests.post(f"{BASE_URL}/v1/voices",
+                      json={"wav_base64": wav_to_base64(wav_path)}, timeout=60)
+    r.raise_for_status()
+    return r.json()["voice_id"]
+
+def synthesize(voice_id, text):
+    r = requests.post(f"{BASE_URL}/v1/speech",
+                      json={"text": text, "voice_id": voice_id,
+                            "cfg_value": 2.0, "inference_timesteps": 10},
+                      timeout=120)
+    r.raise_for_status()
+    return r.content
+
+# ──────────────────────────────────────────────────────────────
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--role", help="只合成指定角色")
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--resume", action="store_true", help="跳过已存在的文件", default=True)
+    args = parser.parse_args()
+
+    lines = LINES
+    if args.role:
+        lines = [(f,r,t) for f,r,t in lines if r == args.role]
+
+    print(f"\n{'='*60}")
+    print(f" TTS v1 — VoxCPM2  ({len(lines)} 条)")
+    print(f" 输出: {OUT_DIR}")
+    print(f"{'='*60}\n")
+
+    if args.dry_run:
+        for f,r,t in lines:
+            out = os.path.join(OUT_DIR, f"{f}_wav_v1.mp3")
+            done = "✓" if os.path.exists(out) else "·"
+            print(f"  [{done}] {f}  [{r}]  {t[:40]}")
+        return
+
+    if not check_server():
+        print(f"[错误] VoxCPM2 服务未响应: {BASE_URL}")
+        sys.exit(1)
+
+    # 预注册音色
+    voice_ids = {}
+    tmp = tempfile.mkdtemp(prefix="voxcpm_")
+    for role_key in set(r for _,r,_ in lines):
+        ref = find_ref_audio(role_key)
+        if not ref:
+            print(f"  [跳过] {role_key}: 未找到参考音频"); continue
+        wav = ref if ref.endswith(".wav") else os.path.join(tmp, f"{role_key}.wav")
+        if not ref.endswith(".wav"):
+            convert_to_wav(ref, wav)
+        print(f"  注册音色: {role_key} ... ", end="", flush=True)
+        vid = register_voice(wav)
+        voice_ids[role_key] = vid
+        print(f"ok ({vid[:8]}...)")
+
+    # 批量合成
+    ok = fail = skip = 0
+    for i, (fname, role, text) in enumerate(lines, 1):
+        out_mp3 = os.path.join(OUT_DIR, f"{fname}_wav_v1.mp3")
+        if args.resume and os.path.exists(out_mp3):
+            print(f"  [{i:3d}/{len(lines)}] SKIP {fname}")
+            skip += 1; continue
+
+        if role not in voice_ids:
+            print(f"  [{i:3d}/{len(lines)}] MISS {fname} (无音色)")
+            fail += 1; continue
+
+        speed = ROLES[role]["speed"]
+        print(f"  [{i:3d}/{len(lines)}] {fname}  text={text[:30]}... ", end="", flush=True)
+        t0 = time.time()
+        try:
+            wav_bytes = synthesize(voice_ids[role], text)
+            raw_wav = os.path.join(tmp, f"{fname}_raw.wav")
+            spd_wav = os.path.join(tmp, f"{fname}_spd.wav")
+            with open(raw_wav, "wb") as f: f.write(wav_bytes)
+            apply_speed(raw_wav, speed, spd_wav)
+            wav_to_mp3(spd_wav, out_mp3)
+            print(f"ok ({time.time()-t0:.1f}s  {os.path.getsize(out_mp3)//1024}KB)")
+            ok += 1
+        except Exception as e:
+            print(f"FAIL: {e}"); fail += 1
+
+    print(f"\n完成: {ok} ok, {skip} skip, {fail} fail → {OUT_DIR}")
+
+if __name__ == "__main__":
+    main()

audio_1/scripts/02_gen_tts_cosyvoice2.py

@@ -0,0 +1,153 @@
+#!/usr/bin/env python3
+"""
+步骤 2：TTS v2 — CosyVoice2-0.5B 批量生成
+输出目录：audio_1/tts/v2_cosyvoice2/
+
+
+使用方法：
+  conda activate cosyvoice
+  python audio_1/scripts/02_gen_tts_cosyvoice2.py
+  python audio_1/scripts/02_gen_tts_cosyvoice2.py --dry-run
+"""
+import os, sys, glob, json, time, tempfile, subprocess, argparse
+import numpy as np
+
+PROJECT_DIR    = "/home/xsl/blind"
+VOICE_DIR      = f"{PROJECT_DIR}/voice"
+OUT_DIR        = f"{PROJECT_DIR}/audio_1/tts/v2_cosyvoice2"
+REF_TEXTS      = f"{PROJECT_DIR}/audio_1/ref_texts.json"
+MODEL_DIR      = "/home/xsl/tools/CosyVoice/pretrained_models/CosyVoice2-0.5B"
+COSYVOICE_DIR  = "/home/xsl/tools/CosyVoice"
+
+os.makedirs(OUT_DIR, exist_ok=True)
+
+sys.path.insert(0, COSYVOICE_DIR)
+sys.path.insert(0, f"{PROJECT_DIR}/audio")
+from batch_tts_voxcpm import ROLES, LINES
+
+with open(REF_TEXTS, encoding="utf-8") as f:
+    REF_TEXT_MAP = json.load(f)
+
+# ──────────────────────────────────────────────────────────────
+def find_ref_audio(role_key):
+    role_dir = os.path.join(VOICE_DIR, ROLES[role_key]["dir"])
+    for ext in ("*.wav", "*.mp3", "*.flac"):
+        files = sorted(glob.glob(os.path.join(role_dir, ext)))
+        if files:
+            return files[0]
+    return None
+
+def load_audio_16k(path):
+    """加载音频，重采样到 16kHz，返回 numpy float32 array"""
+    import librosa
+    audio, _ = librosa.load(path, sr=16000, mono=True)
+    return audio
+
+def wav_to_mp3(src, dst):
+    subprocess.run(["ffmpeg", "-y", "-i", src, "-codec:a", "libmp3lame",
+                    "-b:a", "128k", "-ar", "24000", "-ac", "1", dst],
+                   check=True, capture_output=True)
+
+def save_wav(audio_tensor, sr, path):
+    """保存 torch tensor 或 numpy array 到 WAV"""
+    import soundfile as sf
+    if hasattr(audio_tensor, 'numpy'):
+        audio = audio_tensor.squeeze().numpy()
+    else:
+        audio = np.squeeze(audio_tensor)
+    sf.write(path, audio, sr)
+
+# ──────────────────────────────────────────────────────────────
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--role", help="只合成指定角色")
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--resume", action="store_true", default=True)
+    args = parser.parse_args()
+
+    lines = LINES
+    if args.role:
+        lines = [(f,r,t) for f,r,t in lines if r == args.role]
+
+    print(f"\n{'='*60}")
+    print(f" TTS v2 — CosyVoice2-0.5B  ({len(lines)} 条)")
+    print(f" 输出: {OUT_DIR}")
+    print(f"{'='*60}\n")
+
+    # 预检参考音频
+    ref_cache = {}
+    for role_key in ROLES:
+        ref_path = find_ref_audio(role_key)
+        if ref_path:
+            ref_cache[role_key] = {
+                "path": ref_path,
+                "audio16k": None,  # 懒加载
+                "text": REF_TEXT_MAP.get(role_key, ""),
+            }
+
+    if args.dry_run:
+        for f,r,t in lines:
+            out = os.path.join(OUT_DIR, f"{f}_wav_v1.mp3")
+            done = "✓" if os.path.exists(out) else "·"
+            ref_ok = "✓" if r in ref_cache else "✗"
+            print(f"  [{done}] ref={ref_ok} {f}  [{r}]  {t[:40]}")
+        return
+
+    # 加载模型
+    print("加载 CosyVoice2 模型...", end=" ", flush=True)
+    t0 = time.time()
+    from cosyvoice.cli.cosyvoice import CosyVoice2
+    cosyvoice = CosyVoice2(MODEL_DIR)
+    print(f"就绪 ({time.time()-t0:.1f}s)  SR={cosyvoice.sample_rate}")
+
+    tmp = tempfile.mkdtemp(prefix="cosyvoice2_")
+    ok = fail = skip = 0
+
+    for i, (fname, role, text) in enumerate(lines, 1):
+        out_mp3 = os.path.join(OUT_DIR, f"{fname}_wav_v1.mp3")
+        if args.resume and os.path.exists(out_mp3):
+            print(f"  [{i:3d}/{len(lines)}] SKIP {fname}")
+            skip += 1; continue
+
+        if role not in ref_cache:
+            print(f"  [{i:3d}/{len(lines)}] MISS {fname} (无参考音频)")
+            fail += 1; continue
+
+        speed = ROLES[role]["speed"]
+        ref = ref_cache[role]
+        print(f"  [{i:3d}/{len(lines)}] {fname}  [{role}]  {text[:30]}... ", end="", flush=True)
+        t0 = time.time()
+
+        try:
+            # 懒加载参考音频
+            if ref["audio16k"] is None:
+                ref["audio16k"] = load_audio_16k(ref["path"])
+
+            # 合成：zero-shot 语音克隆
+            import torch
+            prompt_speech = torch.from_numpy(ref["audio16k"]).unsqueeze(0)
+            audio_chunks = []
+            for result in cosyvoice.inference_zero_shot(
+                tts_text=text,
+                prompt_text=ref["text"],
+                prompt_speech_16k=prompt_speech,
+                speed=speed,
+                stream=False,
+            ):
+                audio_chunks.append(result['tts_speech'].squeeze().numpy())
+
+            # 合并并保存
+            full_audio = np.concatenate(audio_chunks) if audio_chunks else np.array([])
+            out_wav = os.path.join(tmp, f"{fname}.wav")
+            import soundfile as sf
+            sf.write(out_wav, full_audio, cosyvoice.sample_rate)
+            wav_to_mp3(out_wav, out_mp3)
+            print(f"ok ({time.time()-t0:.1f}s  {os.path.getsize(out_mp3)//1024}KB)")
+            ok += 1
+        except Exception as e:
+            print(f"FAIL: {e}"); fail += 1
+
+    print(f"\n完成: {ok} ok, {skip} skip, {fail} fail → {OUT_DIR}")
+
+if __name__ == "__main__":
+    main()

audio_1/scripts/04_gen_music_acestep.py

@@ -0,0 +1,122 @@
+#!/usr/bin/env python3
+"""
+步骤 4：音乐 v1 — ACE-Step 1.5 批量生成
+输出目录：audio_1/music/v1_acestep/{music,sfx,ambience}/
+
+使用方法：
+  conda activate acestep
+  python audio_1/scripts/04_gen_music_acestep.py [--dry-run]
+"""
+import os, sys, time, argparse, subprocess
+
+import torch
+torch.backends.cuda.preferred_blas_library("cublaslt")
+
+PROJECT_DIR = "/home/xsl/blind"
+REPO_DIR    = "/home/xsl/tools/ACE-Step-1.5"
+OUT_MUSIC   = f"{PROJECT_DIR}/audio_1/music/v1_acestep/music"
+OUT_SFX     = f"{PROJECT_DIR}/audio_1/music/v1_acestep/sfx"
+OUT_AMB     = f"{PROJECT_DIR}/audio_1/music/v1_acestep/ambience"
+
+for d in (OUT_MUSIC, OUT_SFX, OUT_AMB):
+    os.makedirs(d, exist_ok=True)
+
+sys.path.insert(0, REPO_DIR)
+sys.path.insert(0, f"{PROJECT_DIR}/audio")
+
+# 复用原脚本中的 TRACKS 和 SFX 定义
+from gen_music_acestep import BGM_TRACKS, RINGTONES
+from gen_sfx_acestep   import SFX_TRACKS, AMBIENCE_TRACKS
+
+def wav_to_mp3(src, dst):
+    subprocess.run(["ffmpeg", "-y", "-i", src, "-codec:a", "libmp3lame",
+                    "-b:a", "128k", "-ar", "44100", "-ac", "2", dst],
+                   check=True, capture_output=True)
+
+# ──────────────────────────────────────────────────────────────
+def load_handler():
+    from acestep.handler import AceStepHandler
+    print("加载 ACE-Step 模型...", end=" ", flush=True)
+    t0 = time.time()
+    handler = AceStepHandler()
+    handler.initialize_service(project_root=REPO_DIR, config_path="acestep-v15-turbo", device="cuda")
+    print(f"就绪 ({time.time()-t0:.1f}s)")
+    return handler
+
+def generate_one(handler, track, out_dir):
+    from acestep.inference import GenerationParams, GenerationConfig, generate_music
+    filename = track["filename"]
+    out_wav  = os.path.join(out_dir, f"{filename}_v1.wav")
+    out_mp3  = os.path.join(out_dir, f"{filename}_v1.mp3")
+
+    if os.path.exists(out_mp3):
+        print(f"  [SKIP] {filename}")
+        return "skip"
+
+    note = track.get("note", "")
+    print(f"  [{filename}]  {track['duration']}s  {note} ...", end=" ", flush=True)
+
+    params = GenerationParams(
+        caption=track["caption"],
+        lyrics=track.get("lyrics", ""),
+        duration=track["duration"],
+        instrumental=True,
+        inference_steps=track.get("steps", 20),
+        seed=track.get("seed", 42),
+    )
+    config = GenerationConfig(batch_size=1, audio_format="wav")
+    t0 = time.time()
+    try:
+        result = generate_music(handler, None, params, config, save_dir=out_dir)
+        elapsed = time.time() - t0
+        if result.success:
+            gen_path = result.audios[0]["path"]
+            if gen_path != out_wav:
+                os.rename(gen_path, out_wav)
+            wav_to_mp3(out_wav, out_mp3)
+            size = os.path.getsize(out_mp3) // 1024
+            print(f"ok {elapsed:.1f}s  {size}KB")
+            return "ok"
+        else:
+            print(f"FAIL: {result.error}"); return "fail"
+    except Exception as e:
+        print(f"FAIL: {e}"); return "fail"
+
+# ──────────────────────────────────────────────────────────────
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--category", choices=["music","sfx","ambience","all"], default="all")
+    args = parser.parse_args()
+
+    all_tasks = []
+    if args.category in ("music", "all"):
+        all_tasks += [(t, OUT_MUSIC) for t in BGM_TRACKS + RINGTONES]
+    if args.category in ("sfx", "all"):
+        all_tasks += [(t, OUT_SFX) for t in SFX_TRACKS]
+    if args.category in ("ambience", "all"):
+        all_tasks += [(t, OUT_AMB) for t in AMBIENCE_TRACKS]
+
+    print(f"\n{'='*60}")
+    print(f" 音乐 v1 — ACE-Step  ({len(all_tasks)} 条)")
+    print(f"{'='*60}\n")
+
+    if args.dry_run:
+        for t, d in all_tasks:
+            mp3 = os.path.join(d, f"{t['filename']}_v1.mp3")
+            done = "✓" if os.path.exists(mp3) else "·"
+            print(f"  [{done}] {t['filename']}  {t['duration']}s  → {os.path.basename(d)}/")
+        return
+
+    handler = load_handler()
+    ok = fail = skip = 0
+    for track, out_dir in all_tasks:
+        r = generate_one(handler, track, out_dir)
+        if r == "ok": ok += 1
+        elif r == "skip": skip += 1
+        else: fail += 1
+
+    print(f"\n完成: {ok} ok, {skip} skip, {fail} fail")
+
+if __name__ == "__main__":
+    main()

audio_1/scripts/05_gen_music_stableaudio.py

@@ -0,0 +1,277 @@
+#!/usr/bin/env python3
+"""
+步骤 5：音乐 v2 — Stable Audio Open 1.0 批量生成
+输出目录：audio_1/music/v2_stableaudio/{music,sfx,ambience}/
+
+特点：44.1kHz 立体声，最长 47 秒，音效/氛围质量高，BGM 风格偏弥散
+
+使用方法：
+  conda activate stableaudio
+  python audio_1/scripts/05_gen_music_stableaudio.py [--dry-run]
+"""
+import os, sys, time, subprocess, argparse
+import torch, json
+
+PROJECT_DIR  = "/home/xsl/blind"
+TOOLS_DIR    = "/home/xsl/tools/stable-audio-tools"
+MODEL_CONFIG = "/home/xsl/models/stable-audio-open/model_config.json"
+MODEL_CKPT   = "/home/xsl/models/stable-audio-open/model.safetensors"
+OUT_MUSIC    = f"{PROJECT_DIR}/audio_1/music/v2_stableaudio/music"
+OUT_SFX      = f"{PROJECT_DIR}/audio_1/music/v2_stableaudio/sfx"
+OUT_AMB      = f"{PROJECT_DIR}/audio_1/music/v2_stableaudio/ambience"
+
+for d in (OUT_MUSIC, OUT_SFX, OUT_AMB):
+    os.makedirs(d, exist_ok=True)
+
+sys.path.insert(0, TOOLS_DIR)
+
+# ── 音乐 BGM（Stable Audio 最长 47s，生成片段）─────────────────
+BGM_TRACKS = [
+    {
+        "filename": "lv11_bgm_m1_main_loop",
+        "duration": 47,
+        "prompt": (
+            "lo-fi midnight bedroom, soft piano, light vinyl crackle, "
+            "muted piano, subtle synth pad, occasional distant rain, "
+            "melancholy but warm, alone but not lonely, ambient, no drums, 70 bpm"
+        ),
+        "steps": 200,
+        "seed": 2103,
+        "note": "M1 主题",
+    },
+    {
+        "filename": "lv11_bgm_m2_suspense",
+        "duration": 30,
+        "prompt": (
+            "subtle ambient suspense, cinematic tension, "
+            "low drone, single hanging piano note, sparse strings, "
+            "unsettling but not horror, builds slowly, unresolved chord, 60 bpm"
+        ),
+        "steps": 200,
+        "seed": 2230,
+        "note": "M2 悬疑",
+    },
+    {
+        "filename": "lv11_bgm_m3_zhounan",
+        "duration": 47,
+        "prompt": (
+            "late night lo-fi, warm electric guitar fingerpicking, "
+            "light reverb, distant city sounds, nostalgic, gentle, 65 bpm"
+        ),
+        "steps": 200,
+        "seed": 2345,
+        "note": "M3 周南",
+    },
+    {
+        "filename": "lv11_bgm_m4_ending",
+        "duration": 47,
+        "prompt": (
+            "gentle piano dawn, morning light, hopeful, sparse chords, "
+            "soft string pad, peaceful resolution, 60 bpm"
+        ),
+        "steps": 200,
+        "seed": 2400,
+        "note": "M4 结局",
+    },
+]
+
+# ── 铃声（短促，~10-15s）────────────────────────────────────────
+RINGTONES = [
+    {"filename": "lv11_ring_mom",      "duration": 12, "seed": 3001, "steps": 150,
+     "prompt": "warm marimba ringtone, simple 4-note motif, old Nokia style, 70 bpm, loop, short"},
+    {"filename": "lv11_ring_azhe",     "duration": 12, "seed": 3002, "steps": 150,
+     "prompt": "cold minimal electronic ringtone, 2 notes only, distant, 80 bpm, loop, short"},
+    {"filename": "lv11_ring_xiaomei",  "duration": 12, "seed": 3003, "steps": 150,
+     "prompt": "cheerful pop ringtone, ukulele plucks, 4-bar phrase, energetic, 100 bpm, loop"},
+    {"filename": "lv11_ring_zhounan",  "duration": 12, "seed": 3004, "steps": 150,
+     "prompt": "acoustic guitar single-string ringtone, contemplative, 75 bpm, loop, short"},
+    {"filename": "lv11_ring_hr",       "duration": 12, "seed": 3005, "steps": 150,
+     "prompt": "corporate professional ringtone, neutral marimba, clean, 90 bpm, loop"},
+    {"filename": "lv11_ring_anan",     "duration": 12, "seed": 3006, "steps": 150,
+     "prompt": "fun upbeat ringtone, synth pop, bright, 110 bpm, catchy, loop"},
+    {"filename": "lv11_ring_mom",      "duration": 12, "seed": 3007, "steps": 150,
+     "prompt": "delivery notification ringtone, simple beep melody, neutral, 85 bpm"},
+    {"filename": "lv11_ring_delivery", "duration": 12, "seed": 3008, "steps": 150,
+     "prompt": "delivery notification ringtone, simple beep melody, neutral, 85 bpm"},
+    {"filename": "lv11_ring_unknown",  "duration": 12, "seed": 3009, "steps": 150,
+     "prompt": "eerie unknown caller ringtone, distorted sine wave, unsettling, 60 bpm"},
+    {"filename": "lv11_ring_dorm_group","duration": 12, "seed": 3010, "steps": 150,
+     "prompt": "group chat notification sound, light percussion, cheerful, 95 bpm, loop"},
+    {"filename": "lv11_ring_zhounan",  "duration": 12, "seed": 3011, "steps": 150,
+     "prompt": "acoustic guitar single-string ringtone, contemplative, 75 bpm, loop"},
+]
+
+# ── 音效 SFX ───────────────────────────────────────────────────
+SFX_TRACKS = [
+    {"filename": "sfx_transition_chapter", "duration": 8, "seed": 4001, "steps": 200,
+     "prompt": "cinematic transition swell, soft orchestral rise and fall, gentle fade, film score"},
+    {"filename": "sfx_heartbeat_fast",     "duration": 8, "seed": 4002, "steps": 200,
+     "prompt": "human heartbeat accelerating, close microphone, thumping, tension, anxiety, realistic"},
+    {"filename": "sfx_deep_breath",        "duration": 6, "seed": 4003, "steps": 200,
+     "prompt": "single deep breath, calm, close microphone, soft exhale, meditative, realistic foley"},
+    {"filename": "sfx_rain_heavier",       "duration": 8, "seed": 4004, "steps": 200,
+     "prompt": "rain suddenly intensifying on window glass, urban rain sound, dramatic, realistic"},
+    {"filename": "sfx_rain_lighter",       "duration": 8, "seed": 4005, "steps": 200,
+     "prompt": "light rain on window, soft pattering, urban night, gentle, calming, realistic"},
+    {"filename": "sfx_glass_shatter_short","duration": 5, "seed": 4006, "steps": 200,
+     "prompt": "glass breaking, short, sharp, immediate, foley sound effect"},
+    {"filename": "sfx_door_knock_gentle",  "duration": 5, "seed": 4007, "steps": 200,
+     "prompt": "gentle knocking on door, two light knocks, quiet, hesitant, foley"},
+    {"filename": "sfx_door_knock_firm",    "duration": 5, "seed": 4008, "steps": 200,
+     "prompt": "firm confident door knock, three knocks, hollow door resonance, foley"},
+    {"filename": "sfx_phone_vibrate",      "duration": 5, "seed": 4009, "steps": 200,
+     "prompt": "smartphone vibrating on hard surface, buzzing notification, realistic"},
+    {"filename": "sfx_keyboard_fast",      "duration": 8, "seed": 4010, "steps": 200,
+     "prompt": "fast typing on mechanical keyboard, rapid keystrokes, focused, realistic"},
+    {"filename": "sfx_keyboard_slow",      "duration": 8, "seed": 4011, "steps": 200,
+     "prompt": "slow hesitant typing, single key presses, pauses, uncertain typing, realistic"},
+    {"filename": "sfx_kettle_whistle",     "duration": 8, "seed": 4012, "steps": 200,
+     "prompt": "electric kettle boiling and whistling, kitchen, steam, morning, realistic foley"},
+    {"filename": "sfx_water_boiling",      "duration": 8, "seed": 4013, "steps": 200,
+     "prompt": "water boiling in pot, bubbling, kitchen ambience, realistic foley sound"},
+    {"filename": "sfx_paper_turning",      "duration": 5, "seed": 4014, "steps": 200,
+     "prompt": "paper page turning, subtle rustling, quiet, realistic close-mic foley"},
+    {"filename": "sfx_night_bus",          "duration": 10, "seed": 4015, "steps": 200,
+     "prompt": "night bus interior sound, engine rumble, gentle sway, city at night, ambient"},
+    {"filename": "sfx_cat_yowl",           "duration": 5, "seed": 4016, "steps": 200,
+     "prompt": "cat yowling outside at night, distant, urban night sound, realistic"},
+    {"filename": "sfx_ending_chime",       "duration": 6, "seed": 4017, "steps": 200,
+     "prompt": "soft ending chime, single bell tone, resonant, gentle, conclusive, musical"},
+    {"filename": "sfx_breathing_unknown",  "duration": 8, "seed": 4018, "steps": 200,
+     "prompt": "quiet breathing, slightly nervous, close microphone, intimate, human, realistic"},
+    {"filename": "sfx_deep_breath",        "duration": 6, "seed": 4019, "steps": 200,
+     "prompt": "slow deep calming breath, inhale and exhale, meditation, close mic, peaceful"},
+]
+
+# ── 环境氛围 ──────────────────────────────────────────────────
+AMBIENCE_TRACKS = [
+    {"filename": "amb_apartment_night_01", "duration": 47, "seed": 5001, "steps": 200,
+     "prompt": "apartment at night, city sounds, distant traffic, quiet hum, urban isolation, ambient"},
+    {"filename": "amb_apartment_night_02", "duration": 47, "seed": 5002, "steps": 200,
+     "prompt": "late night apartment, occasional car passing, silence, urban night, subtle AC hum"},
+    {"filename": "amb_bar_loud_01",        "duration": 30, "seed": 5003, "steps": 200,
+     "prompt": "busy bar ambience, crowd chatter, glasses clinking, background music, lively, indoor"},
+    {"filename": "amb_rain_window_01",     "duration": 47, "seed": 5004, "steps": 200,
+     "prompt": "rain on window glass, cozy indoor, distant thunder rumble, peaceful, continuous"},
+    {"filename": "amb_kitchen_morning",    "duration": 30, "seed": 5005, "steps": 200,
+     "prompt": "morning kitchen sounds, birds outside, soft light, gentle hum, peaceful domestic"},
+    {"filename": "amb_dawn_birds",         "duration": 47, "seed": 5006, "steps": 200,
+     "prompt": "dawn birdsong, early morning outdoor, multiple bird species, peaceful, hopeful"},
+]
+
+# ──────────────────────────────────────────────────────────────
+def wav_to_mp3(src, dst):
+    subprocess.run(["ffmpeg", "-y", "-i", src, "-codec:a", "libmp3lame",
+                    "-b:a", "128k", "-ar", "44100", "-ac", "2", dst],
+                   check=True, capture_output=True)
+
+def generate_one(track, out_dir, sa_model, sr):
+    """使用 generate_diffusion_cond 直接生成音频"""
+    import numpy as np
+    from einops import rearrange
+    from stable_audio_tools.inference.generation import generate_diffusion_cond
+    import torchaudio
+
+    filename = track["filename"]
+    out_wav   = os.path.join(out_dir, f"{filename}_v1.wav")
+    out_mp3   = os.path.join(out_dir, f"{filename}_v1.mp3")
+
+    if os.path.exists(out_mp3):
+        print(f"  [SKIP] {filename}")
+        return "skip"
+
+    note = track.get("note", "")
+    duration = min(track["duration"], 47)
+    print(f"  [{filename}]  {duration}s  {note}... ", end="", flush=True)
+    t0 = time.time()
+
+    try:
+        device = next(sa_model.parameters()).device
+        conditioning = [{"prompt": track["prompt"], "seconds_start": 0, "seconds_total": duration}]
+        audio = generate_diffusion_cond(
+            model=sa_model,
+            conditioning=conditioning,
+            steps=track.get("steps", 200),
+            cfg_scale=7.0,
+            batch_size=1,
+            sample_size=sr * duration,
+            seed=track.get("seed", 42),
+            device=device,
+            sampler_type="dpmpp-3m-sde",
+            sigma_min=0.03,
+            sigma_max=1000,
+        )
+        # audio shape: [batch, channels, samples]
+        audio = audio[:, :, :duration * sr]
+        audio = rearrange(audio, "b d n -> d (b n)")
+        audio = audio.to(torch.float32)
+        peak = torch.max(torch.abs(audio))
+        if peak > 0:
+            audio = audio / peak
+        audio = audio.clamp(-1, 1).mul(32767).to(torch.int16).cpu()
+        torchaudio.save(out_wav, audio, sr)
+        wav_to_mp3(out_wav, out_mp3)
+        size = os.path.getsize(out_mp3) // 1024
+        print(f"ok {time.time()-t0:.1f}s  {size}KB")
+        return "ok"
+    except Exception as e:
+        import traceback; traceback.print_exc()
+        print(f"FAIL: {e}"); return "fail"
+
+# ──────────────────────────────────────────────────────────────
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--dry-run", action="store_true")
+    parser.add_argument("--category", choices=["music","sfx","ambience","all"], default="all")
+    args = parser.parse_args()
+
+    all_tasks = []
+    if args.category in ("music", "all"):
+        all_tasks += [(t, OUT_MUSIC) for t in BGM_TRACKS + RINGTONES]
+    if args.category in ("sfx", "all"):
+        all_tasks += [(t, OUT_SFX) for t in SFX_TRACKS]
+    if args.category in ("ambience", "all"):
+        all_tasks += [(t, OUT_AMB) for t in AMBIENCE_TRACKS]
+
+    # 去重（RINGTONES 里有重复 filename）
+    seen = set()
+    deduped = []
+    for t, d in all_tasks:
+        key = (t["filename"], d)
+        if key not in seen:
+            seen.add(key)
+            deduped.append((t, d))
+    all_tasks = deduped
+
+    print(f"\n{'='*60}")
+    print(f" 音乐 v2 — Stable Audio Open 1.0  ({len(all_tasks)} 条)")
+    print(f"{'='*60}\n")
+
+    if args.dry_run:
+        for t, d in all_tasks:
+            mp3 = os.path.join(d, f"{t['filename']}_v1.mp3")
+            done = "✓" if os.path.exists(mp3) else "·"
+            print(f"  [{done}] {t['filename']}  {t['duration']}s  → {os.path.basename(d)}/")
+        return
+
+    # 加载模型
+    print(f"加载 Stable Audio 模型...", end=" ", flush=True)
+    t0 = time.time()
+    from stable_audio_tools.interface.gradio import load_model as sa_load_model
+    import json
+    with open(MODEL_CONFIG) as f:
+        model_config = json.load(f)
+    sa_model, _ = sa_load_model(model_config=model_config, model_ckpt_path=MODEL_CKPT, device="cuda", model_half=False)
+    print(f"就绪 ({time.time()-t0:.1f}s)")
+
+    ok = fail = skip = 0
+    for track, out_dir in all_tasks:
+        r = generate_one(track, out_dir, sa_model, model_config["sample_rate"])
+        if r == "ok": ok += 1
+        elif r == "skip": skip += 1
+        else: fail += 1
+
+    print(f"\n完成: {ok} ok, {skip} skip, {fail} fail")
+
+if __name__ == "__main__":
+    main()

audio_1/scripts/06_convert_all_to_mp3.py

@@ -0,0 +1,83 @@
+#!/usr/bin/env python3
+"""
+步骤 6：格式整理 — 将 audio_1 目录下所有残留 WAV 转换为 MP3，并清理中间文件
+输出：各子目录内的 *_v1.mp3
+
+使用方法：
+  python audio_1/scripts/06_convert_all_to_mp3.py [--dry-run] [--clean-wav]
+"""
+import os, subprocess, argparse
+
+AUDIO_1_DIR = "/home/xsl/blind/audio_1"
+
+def wav_to_mp3(src, dst):
+    subprocess.run(["ffmpeg", "-y", "-i", src, "-codec:a", "libmp3lame",
+                    "-b:a", "128k", "-ar", "44100", "-ac", "2", dst],
+                   check=True, capture_output=True)
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--dry-run", action="store_true", help="只列出，不执行")
+    parser.add_argument("--clean-wav", action="store_true", help="转换后删除 WAV 源文件")
+    args = parser.parse_args()
+
+    wavs_found = []
+    for root, dirs, files in os.walk(AUDIO_1_DIR):
+        # 跳过 scripts 目录
+        if "scripts" in root:
+            continue
+        for fname in files:
+            if fname.endswith(".wav"):
+                wav_path = os.path.join(root, fname)
+                mp3_path = wav_path.replace(".wav", ".mp3")
+                wavs_found.append((wav_path, mp3_path))
+
+    if not wavs_found:
+        print("无需转换的 WAV 文件。")
+        return
+
+    print(f"\n找到 {len(wavs_found)} 个 WAV 文件\n")
+    ok = skip = fail = 0
+    for wav, mp3 in wavs_found:
+        rel = os.path.relpath(wav, AUDIO_1_DIR)
+        if os.path.exists(mp3):
+            print(f"  [SKIP] {rel}")
+            skip += 1
+            continue
+        if args.dry_run:
+            print(f"  [DRY]  {rel} → {os.path.basename(mp3)}")
+            continue
+        try:
+            wav_to_mp3(wav, mp3)
+            size = os.path.getsize(mp3) // 1024
+            print(f"  [OK]   {rel} → {size}KB")
+            if args.clean_wav:
+                os.remove(wav)
+            ok += 1
+        except Exception as e:
+            print(f"  [FAIL] {rel}: {e}")
+            fail += 1
+
+    print(f"\n完成: {ok} ok, {skip} skip, {fail} fail")
+
+    # 统计各目录 MP3 数量
+    print("\n─── 音频资源统计 ───")
+    subdirs = [
+        ("tts/v1_voxcpm2",        "TTS v1 VoxCPM2"),
+        ("music/v1_acestep/music",    "Music v1 ACE-Step BGM"),
+        ("music/v1_acestep/sfx",      "Music v1 ACE-Step SFX"),
+        ("music/v1_acestep/ambience", "Music v1 ACE-Step Amb"),
+        ("music/v2_stableaudio/music",    "Music v2 StableAudio BGM"),
+        ("music/v2_stableaudio/sfx",      "Music v2 StableAudio SFX"),
+        ("music/v2_stableaudio/ambience", "Music v2 StableAudio Amb"),
+    ]
+    for rel_dir, label in subdirs:
+        full = os.path.join(AUDIO_1_DIR, rel_dir)
+        if not os.path.isdir(full):
+            count = 0
+        else:
+            count = len([f for f in os.listdir(full) if f.endswith(".mp3")])
+        print(f"  {label:<30} {count:>4} 条")
+
+if __name__ == "__main__":
+    main()