#!/usr/bin/env python3
"""批次聲音轉換(Seed-VC):把一批語音(例如曉臻念的旁白 mp3)轉成目標講師的音色。
用法(在 tools/voices/clone/seed-vc 目錄下以 sv-venv 的 python 執行):python ../vc-batch.py < job.json
job.json:{"target": "<目標聲音 wav>", "steps": 25, "cfg": 0.7, "items": [{"src": "a.mp3", "out": "a.wav"}, ...]}
模型與目標聲音特徵只載入/計算一次;已存在的 out 會跳過。轉換內容與 seed-vc/inference.py 相同(不使用 f0 條件)。
"""
import json, os, sys, time, types
import numpy as np
sys.path.insert(0, os.getcwd())
import torch, torchaudio, librosa
import inference as sv

job = json.load(sys.stdin)
torch.set_num_threads(int(os.environ.get('VC_THREADS', '12')))
args = types.SimpleNamespace(f0_condition=False, checkpoint=None, config=None, fp16=False)
model, semantic_fn, f0_fn, vocoder_fn, campplus_model, mel_fn, mel_fn_args = sv.load_models(args)
device = sv.device
sr = 22050; hop = 256
max_ctx = sr // hop * 30; ovl_f = 16; ovl_w = ovl_f * hop
steps = int(job.get('steps', 25)); cfg = float(job.get('cfg', 0.7))

with torch.no_grad():
    ref = librosa.load(job['target'], sr=mel_fn_args['sampling_rate'])[0]
    ref = torch.tensor(ref[:sr * 25]).unsqueeze(0).float().to(device)
    ref16 = torchaudio.functional.resample(ref, sr, 16000)
    S_ori = semantic_fn(ref16)
    mel2 = mel_fn(ref.float())
    feat2 = torchaudio.compliance.kaldi.fbank(ref16, num_mel_bins=80, dither=0, sample_frequency=16000)
    feat2 = feat2 - feat2.mean(dim=0, keepdim=True)
    style2 = campplus_model(feat2.unsqueeze(0))
    prompt_cond = model.length_regulator(S_ori, ylens=torch.LongTensor([mel2.size(2)]).to(device), n_quantizers=3, f0=None)[0]

def convert(path):
    src = librosa.load(path, sr=mel_fn_args['sampling_rate'])[0]
    src = torch.tensor(src).unsqueeze(0).float().to(device)
    w16 = torchaudio.functional.resample(src, sr, 16000)
    if w16.size(-1) <= 16000 * 30:
        S_alt = semantic_fn(w16)
    else:
        parts, buf, t = [], None, 0
        while t < w16.size(-1):
            chunk = w16[:, t:t + 16000 * 30] if buf is None else torch.cat([buf, w16[:, t:t + 16000 * 25]], dim=-1)
            S = semantic_fn(chunk)
            parts.append(S if t == 0 else S[:, 50 * 5:])
            buf = chunk[:, -16000 * 5:]
            t += 30 * 16000 if t == 0 else chunk.size(-1) - 16000 * 5
        S_alt = torch.cat(parts, dim=1)
    mel = mel_fn(src.float())
    cond = model.length_regulator(S_alt, ylens=torch.LongTensor([mel.size(2)]).to(device), n_quantizers=3, f0=None)[0]
    max_src = max_ctx - mel2.size(2)
    done, chunks, prev = 0, [], None
    while done < cond.size(1):
        cc = cond[:, done:done + max_src]
        last = done + max_src >= cond.size(1)
        cat = torch.cat([prompt_cond, cc], dim=1)
        vt = model.cfm.inference(cat, torch.LongTensor([cat.size(1)]).to(device), mel2, style2, None, steps, inference_cfg_rate=cfg)
        vt = vt[:, :, mel2.size(-1):]
        wave = vocoder_fn(vt.float()).squeeze()[None, :]
        if done == 0:
            if last:
                chunks.append(wave[0].cpu().numpy()); break
            chunks.append(wave[0, :-ovl_w].cpu().numpy()); prev = wave[0, -ovl_w:]; done += vt.size(2) - ovl_f
        elif last:
            chunks.append(sv.crossfade(prev.cpu().numpy(), wave[0].cpu().numpy(), ovl_w)); break
        else:
            chunks.append(sv.crossfade(prev.cpu().numpy(), wave[0, :-ovl_w].cpu().numpy(), ovl_w)); prev = wave[0, -ovl_w:]; done += vt.size(2) - ovl_f
    return torch.tensor(np.concatenate(chunks))[None, :].float()

n = len(job['items']); t0 = time.time(); audio_s = 0.0
for k, it in enumerate(job['items']):
    if os.path.exists(it['out']) and os.path.getsize(it['out']) > 1000:
        continue
    with torch.no_grad():
        w = convert(it['src'])
    tmp = it['out'] + '.part.wav'
    torchaudio.save(tmp, w, sr); os.replace(tmp, it['out'])
    audio_s += w.size(-1) / sr
    print(f"[{k + 1}/{n}] {audio_s:.0f}s 音訊,累計 {time.time() - t0:.0f}s,RTF {(time.time() - t0) / max(audio_s, 1):.1f}", file=sys.stderr, flush=True)
print('VC_DONE', file=sys.stderr)
