Add tts skill

SKILL.md

@@ -0,0 +1,25 @@
+---
+name: MockingBird
+description: AI voice cloning and TTS toolkit. The only active skill is `speak`.
+---
+
+# MockingBird
+
+Voice cloning / text-to-speech toolkit built on encoder-synthesizer-vocoder pipeline.
+
+## Available Skills
+
+| Skill | Path | Purpose |
+|-------|------|---------|
+| speak | `skills/speak/SKILL.md` | Text-to-speech (Kokoro local / Noiz cloud), timeline rendering, voice cloning |
+
+## Project Layout
+
+- `models/` — encoder, synthesizer, vocoder, ppg2mel, ppg_extractor
+- `control/` — CLI and GUI entry points (toolbox, mkgui)
+- `skills/speak/` — TTS skill scripts (`tts.sh`, `noiz_tts.py`, `render_timeline.py`, `text_to_srt.py`)
+- `data/ckpt/` — pretrained checkpoints
+
+## When to Use
+
+If the user asks for anything related to speech synthesis, TTS, voice cloning, dubbing, or audio generation, read and follow `skills/speak/SKILL.md`.

skills/speak/SKILL.md

@@ -0,0 +1,114 @@
+---
+name: speak
+description: Convert text into speech with Kokoro or Noiz, including simple and timeline-aligned modes.
+---
+
+# speak
+
+Convert any text into speech audio. Supports two backends (Kokoro local, Noiz cloud), two modes (simple or timeline-accurate), and per-segment voice control.
+
+## Triggers
+
+- text to speech / speak / say / tts
+- voice clone / dubbing 
+- epub to audio / srt to audio / convert to audio
+
+## Simple Mode — text to audio
+
+```bash
+# Kokoro (auto-detected when installed)
+bash skills/speak/scripts/tts.sh speak -t "Hello world" -v af_sarah -o hello.wav
+bash skills/speak/scripts/tts.sh speak -f article.txt -v zf_xiaoni --lang cmn -o out.mp3 --format mp3
+
+# Noiz (auto-detected when NOIZ_API_KEY is set, or force with --backend noiz)
+# If --voice-id is omitted, the script prints 5 available built-in voices and exits.
+# Pick one from the output and re-run with --voice-id <id>.
+bash skills/speak/scripts/tts.sh speak -f input.txt --voice-id voice_abc --auto-emotion --emo '{"Joy":0.5}' -o out.wav
+
+# Noiz: optional --duration (float, seconds, range (0, 36]) for target audio length
+bash skills/speak/scripts/tts.sh speak -t "Short line" --voice-id voice_abc --duration 3.5 -o out.wav
+
+# Voice cloning (Noiz only — no voice-id needed, uses ref audio)
+# Use your own reference audio: local file path or URL (only when using Noiz).
+bash skills/speak/scripts/tts.sh speak -t "Hello" --ref-audio ./ref.wav -o clone.wav
+bash skills/speak/scripts/tts.sh speak -t "Hello" --ref-audio https://example.com/my_voice.wav -o clone.wav
+```
+
+## Timeline Mode — SRT to time-aligned audio
+
+For precise per-segment timing (dubbing, subtitles, video narration).
+
+### Step 1: Get or create an SRT
+
+If the user doesn't have one, generate from text:
+
+```bash
+bash skills/speak/scripts/tts.sh to-srt -i article.txt -o article.srt
+bash skills/speak/scripts/tts.sh to-srt -i article.txt -o article.srt --cps 15 --gap 500
+```
+
+`--cps` = characters per second (default 4, good for Chinese; ~15 for English). The agent can also write SRT manually.
+
+### Step 2: Create a voice map
+
+JSON file controlling default + per-segment voice settings. `segments` keys support single index `"3"` or range `"5-8"`.
+
+Kokoro voice map:
+
+```json
+{
+  "default": { "voice": "zf_xiaoni", "lang": "cmn" },
+  "segments": {
+    "1": { "voice": "zm_yunxi" },
+    "5-8": { "voice": "af_sarah", "lang": "en-us", "speed": 0.9 }
+  }
+}
+```
+
+Noiz voice map (adds `emo`, `reference_audio` support). `reference_audio` can be a local path or a URL (user’s own audio; Noiz only):
+
+```json
+{
+  "default": { "voice_id": "voice_123", "target_lang": "zh" },
+  "segments": {
+    "1": { "voice_id": "voice_host", "emo": { "Joy": 0.6 } },
+    "2-4": { "reference_audio": "./refs/guest.wav" }
+  }
+}
+```
+
+**Dynamic Reference Audio Slicing**:
+If you are translating or dubbing a video and want each sentence to automatically use the audio from the original video at the exact same timestamp as its reference audio, use the `--ref-audio-track` argument instead of setting `reference_audio` in the map:
+```bash
+bash skills/speak/scripts/tts.sh render --srt input.srt --voice-map vm.json --ref-audio-track original_video.mp4 -o output.wav
+```
+
+See `examples/` for full samples.
+
+### Step 3: Render
+
+```bash
+bash skills/speak/scripts/tts.sh render --srt input.srt --voice-map vm.json -o output.wav
+bash skills/speak/scripts/tts.sh render --srt input.srt --voice-map vm.json --backend noiz --auto-emotion -o output.wav
+```
+
+## When to Choose Which
+
+| Need | Recommended |
+|------|-------------|
+| Just read text aloud, no fuss | Kokoro (default) |
+| EPUB/PDF audiobook with chapters | Kokoro (native support) |
+| Voice blending (`"v1:60,v2:40"`) | Kokoro |
+| Voice cloning from reference audio | Noiz |
+| Emotion control (`emo` param) | Noiz |
+| Exact server-side duration per segment | Noiz |
+
+> When the user needs emotion control + voice cloning + precise duration together, Noiz is the only backend that supports all three.
+
+## Requirements
+
+- `ffmpeg` in PATH (timeline mode)
+- Noiz: get your API key at [developers.noiz.ai](https://developers.noiz.ai), then run `bash skills/speak/scripts/tts.sh config --set-api-key YOUR_KEY`
+- Kokoro: if already installed, pass `--backend kokoro` to use the local backend
+
+For backend details and full argument reference, see [reference.md](reference.md).

skills/speak/scripts/noiz_tts.py

@@ -0,0 +1,201 @@
+#!/usr/bin/env python3
+"""Simple TTS via Noiz API (no timeline).
+
+Supports direct text or text-file input, optional emotion enhancement,
+voice cloning via reference audio, and emotion parameters.
+Use kokoro-tts CLI directly for the Kokoro backend (no wrapper needed).
+"""
+import argparse
+import base64
+import binascii
+import json
+import sys
+from pathlib import Path
+from typing import Any, Dict, Optional
+
+import requests
+
+
+def normalize_api_key_base64(api_key: str) -> str:
+    key = api_key.strip()
+    if not key:
+        return key
+    padded = key + ("=" * (-len(key) % 4))
+    try:
+        decoded = base64.b64decode(padded, validate=True)
+        canonical = base64.b64encode(decoded).decode("ascii").rstrip("=")
+        if decoded and canonical == key.rstrip("="):
+            return key
+    except binascii.Error:
+        pass
+    return base64.b64encode(key.encode("utf-8")).decode("ascii")
+
+
+def call_emotion_enhance(
+    base_url: str, api_key: str, text: str, timeout: int
+) -> str:
+    resp = requests.post(
+        f"{base_url.rstrip('/')}/emotion-enhance",
+        headers={"Authorization": api_key, "Content-Type": "application/json"},
+        json={"text": text},
+        timeout=timeout,
+    )
+    if resp.status_code != 200:
+        raise RuntimeError(
+            f"/emotion-enhance failed: status={resp.status_code}, body={resp.text}"
+        )
+    enhanced = resp.json().get("data", {}).get("emotion_enhance")
+    if not enhanced:
+        raise RuntimeError(f"/emotion-enhance returned no data: {resp.text}")
+    return enhanced
+
+
+def synthesize(
+    base_url: str,
+    api_key: str,
+    text: str,
+    voice_id: Optional[str],
+    reference_audio: Optional[Path],
+    output_format: str,
+    speed: float,
+    emo: Optional[str],
+    target_lang: Optional[str],
+    similarity_enh: bool,
+    save_voice: bool,
+    duration: Optional[float],
+    timeout: int,
+    out_path: Path,
+) -> float:
+    if duration is not None and not (0 < duration <= 36):
+        raise ValueError("duration must be in range (0, 36] seconds")
+    url = f"{base_url.rstrip('/')}/text-to-speech"
+    data: Dict[str, str] = {
+        "text": text,
+        "output_format": output_format,
+        "speed": str(speed),
+    }
+    if voice_id:
+        data["voice_id"] = voice_id
+    if emo:
+        data["emo"] = emo
+    if target_lang:
+        data["target_lang"] = target_lang
+    if similarity_enh:
+        data["similarity_enh"] = "true"
+    if save_voice:
+        data["save_voice"] = "true"
+    if duration is not None:
+        data["duration"] = str(duration)
+
+    files = None
+    if reference_audio:
+        if not reference_audio.exists():
+            raise FileNotFoundError(f"Reference audio not found: {reference_audio}")
+        files = {
+            "file": (
+                reference_audio.name,
+                reference_audio.open("rb"),
+                "application/octet-stream",
+            )
+        }
+    elif not voice_id:
+        raise ValueError("Either --voice-id or --reference-audio is required.")
+
+    try:
+        resp = requests.post(
+            url,
+            headers={"Authorization": api_key},
+            data=data,
+            files=files,
+            timeout=timeout,
+        )
+    finally:
+        if files and files["file"][1]:
+            files["file"][1].close()
+
+    if resp.status_code != 200:
+        raise RuntimeError(
+            f"/text-to-speech failed: status={resp.status_code}, body={resp.text}"
+        )
+
+    out_path.parent.mkdir(parents=True, exist_ok=True)
+    out_path.write_bytes(resp.content)
+    dur = resp.headers.get("X-Audio-Duration")
+    return float(dur) if dur else -1.0
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(description="Simple TTS via Noiz API (no timeline).")
+    g = parser.add_mutually_exclusive_group(required=True)
+    g.add_argument("--text", help="Text string to synthesize")
+    g.add_argument("--text-file", help="Path to text file")
+    parser.add_argument("--api-key", required=True)
+    parser.add_argument("--voice-id")
+    parser.add_argument("--reference-audio", help="Local audio for voice cloning")
+    parser.add_argument("--output", required=True)
+    parser.add_argument("--base-url", default="https://noiz.ai/v1")
+    parser.add_argument("--output-format", choices=["wav", "mp3"], default="wav")
+    parser.add_argument("--auto-emotion", action="store_true")
+    parser.add_argument("--emo", help='Emotion JSON string, e.g. \'{"Joy":0.5}\'')
+    parser.add_argument("--speed", type=float, default=1.0)
+    parser.add_argument("--target-lang")
+    parser.add_argument("--similarity-enh", action="store_true")
+    parser.add_argument("--save-voice", action="store_true")
+    parser.add_argument(
+        "--duration",
+        type=float,
+        default=None,
+        metavar="SEC",
+        help="Target audio duration in seconds (0, 36], optional",
+    )
+    parser.add_argument("--timeout-sec", type=int, default=120)
+    args = parser.parse_args()
+    args.api_key = normalize_api_key_base64(args.api_key)
+
+    try:
+        if args.text_file:
+            text = Path(args.text_file).read_text(encoding="utf-8").strip()
+        else:
+            text = args.text
+
+        if not text:
+            raise ValueError("Input text is empty.")
+
+        if len(text) > 5000:
+            print(
+                f"Warning: text is {len(text)} chars (max 5000). "
+                "Consider chunking for long texts.",
+                file=sys.stderr,
+            )
+
+        if args.auto_emotion:
+            text = call_emotion_enhance(
+                args.base_url, args.api_key, text, args.timeout_sec
+            )
+
+        ref = Path(args.reference_audio) if args.reference_audio else None
+        out_duration = synthesize(
+            base_url=args.base_url,
+            api_key=args.api_key,
+            text=text,
+            voice_id=args.voice_id,
+            reference_audio=ref,
+            output_format=args.output_format,
+            speed=args.speed,
+            emo=args.emo,
+            target_lang=args.target_lang,
+            similarity_enh=args.similarity_enh,
+            save_voice=args.save_voice,
+            duration=args.duration,
+            timeout=args.timeout_sec,
+            out_path=Path(args.output),
+        )
+        print(f"Done. Output: {args.output} (duration: {out_duration}s)")
+        return 0
+    except Exception as exc:
+        print(f"Error: {exc}", file=sys.stderr)
+        return 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

skills/speak/scripts/render_timeline.py

@@ -0,0 +1,486 @@
+#!/usr/bin/env python3
+"""Timeline mode: render SRT to timeline-accurate audio.
+
+Supports two backends:
+  - kokoro (default): local CLI, uses ffmpeg atempo for duration matching
+  - noiz: cloud API with server-side duration forcing, emotion, voice cloning
+
+Parses SRT, resolves per-segment voice config from a voice-map JSON,
+calls TTS for each segment, normalizes to exact duration, delays to
+correct start time, and mixes into one timeline track.
+"""
+import argparse
+import base64
+import binascii
+import json
+import re
+import shutil
+import subprocess
+import sys
+import tempfile
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Any, Dict, List, Optional, Tuple
+
+TIMESTAMP_RE = re.compile(r"^(\d{2}):(\d{2}):(\d{2})[,.](\d{3})$")
+
+
+def normalize_api_key_base64(api_key: str) -> str:
+    key = api_key.strip()
+    if not key:
+        return key
+    padded = key + ("=" * (-len(key) % 4))
+    try:
+        decoded = base64.b64decode(padded, validate=True)
+        canonical = base64.b64encode(decoded).decode("ascii").rstrip("=")
+        if decoded and canonical == key.rstrip("="):
+            return key
+    except binascii.Error:
+        pass
+    return base64.b64encode(key.encode("utf-8")).decode("ascii")
+
+
+@dataclass
+class Cue:
+    index: int
+    start_ms: int
+    end_ms: int
+    text: str
+
+    @property
+    def duration_ms(self) -> int:
+        return max(1, self.end_ms - self.start_ms)
+
+
+# ── SRT parsing ──────────────────────────────────────────────────────
+
+
+def parse_timestamp_ms(value: str) -> int:
+    match = TIMESTAMP_RE.match(value.strip())
+    if not match:
+        raise ValueError(f"Invalid SRT timestamp: {value}")
+    hh, mm, ss, ms = map(int, match.groups())
+    return ((hh * 60 + mm) * 60 + ss) * 1000 + ms
+
+
+def parse_srt(path: Path) -> List[Cue]:
+    content = path.read_text(encoding="utf-8", errors="replace")
+    blocks = re.split(r"\n\s*\n", content.strip())
+    cues: List[Cue] = []
+    for block in blocks:
+        lines = [ln.rstrip() for ln in block.splitlines() if ln.strip()]
+        if len(lines) < 3:
+            continue
+        try:
+            idx = int(lines[0])
+        except ValueError:
+            continue
+        if "-->" not in lines[1]:
+            continue
+        start_raw, end_raw = [s.strip() for s in lines[1].split("-->", 1)]
+        start_ms = parse_timestamp_ms(start_raw)
+        end_ms = parse_timestamp_ms(end_raw)
+        text = "\n".join(lines[2:]).strip()
+        if text:
+            cues.append(Cue(index=idx, start_ms=start_ms, end_ms=end_ms, text=text))
+    if not cues:
+        raise ValueError("No valid cues parsed from SRT.")
+    return cues
+
+
+# ── Voice map resolution ─────────────────────────────────────────────
+
+
+def parse_segment_key(key: str) -> Tuple[int, int]:
+    key = key.strip()
+    if "-" in key:
+        left, right = key.split("-", 1)
+        return int(left), int(right)
+    v = int(key)
+    return v, v
+
+
+def resolve_segment_cfg(index: int, config: Dict[str, Any]) -> Dict[str, Any]:
+    merged = dict(config.get("default", {}))
+    for key, seg_cfg in config.get("segments", {}).items():
+        lo, hi = parse_segment_key(key)
+        if lo <= index <= hi:
+            merged.update(seg_cfg)
+    return merged
+
+
+# ── ffmpeg helpers ────────────────────────────────────────────────────
+
+
+def _run_ff(cmd: List[str]) -> None:
+    proc = subprocess.run(cmd, capture_output=True, text=True)
+    if proc.returncode != 0:
+        raise RuntimeError(f"ffmpeg failed: {' '.join(cmd)}\n{proc.stderr}")
+
+
+def ensure_ffmpeg() -> None:
+    if not shutil.which("ffmpeg"):
+        raise RuntimeError("ffmpeg not found in PATH.")
+
+
+def probe_duration_ms(path: Path) -> float:
+    proc = subprocess.run(
+        [
+            "ffprobe", "-v", "error", "-show_entries", "format=duration",
+            "-of", "default=noprint_wrappers=1:nokey=1", str(path),
+        ],
+        capture_output=True, text=True,
+    )
+    if proc.returncode != 0:
+        raise RuntimeError(f"ffprobe failed on {path}: {proc.stderr}")
+    return float(proc.stdout.strip()) * 1000
+
+
+def normalize_duration_pad_trim(inp: Path, outp: Path, target_ms: int) -> None:
+    """Pad short audio then trim to exact target duration (Noiz backend)."""
+    sec = target_ms / 1000.0
+    _run_ff([
+        "ffmpeg", "-y", "-i", str(inp),
+        "-af", f"apad=pad_dur={sec:.3f}",
+        "-t", f"{sec:.3f}", str(outp),
+    ])
+
+
+def normalize_duration_atempo(inp: Path, outp: Path, target_ms: int) -> None:
+    """Use atempo to stretch/compress audio to target duration (Kokoro backend)."""
+    actual_ms = probe_duration_ms(inp)
+    if actual_ms <= 0:
+        normalize_duration_pad_trim(inp, outp, target_ms)
+        return
+
+    ratio = actual_ms / target_ms
+    # atempo accepts 0.5–100.0; chain filters for extreme ratios
+    filters = []
+    r = ratio
+    while r > 100.0:
+        filters.append("atempo=100.0")
+        r /= 100.0
+    while r < 0.5:
+        filters.append("atempo=0.5")
+        r /= 0.5
+    filters.append(f"atempo={r:.6f}")
+
+    _run_ff([
+        "ffmpeg", "-y", "-i", str(inp),
+        "-af", ",".join(filters),
+        "-t", f"{target_ms / 1000.0:.3f}", str(outp),
+    ])
+
+
+def delay_segment(inp: Path, outp: Path, start_ms: int) -> None:
+    _run_ff([
+        "ffmpeg", "-y", "-i", str(inp),
+        "-af", f"adelay={start_ms}:all=1", str(outp),
+    ])
+
+
+def mix_all(inputs: List[Path], outp: Path, total_ms: int) -> None:
+    if not inputs:
+        raise ValueError("No segments to mix.")
+    cmd = ["ffmpeg", "-y"]
+    for p in inputs:
+        cmd += ["-i", str(p)]
+    cmd += [
+        "-filter_complex",
+        f"amix=inputs={len(inputs)}:duration=longest:dropout_transition=0",
+        "-t", f"{total_ms / 1000.0:.3f}", str(outp),
+    ]
+    _run_ff(cmd)
+
+
+# ── Noiz backend ─────────────────────────────────────────────────────
+
+
+def _noiz_emotion_enhance(
+    base_url: str, api_key: str, text: str, timeout: int
+) -> str:
+    import requests  # noqa: delayed import so kokoro path doesn't need requests
+
+    resp = requests.post(
+        f"{base_url.rstrip('/')}/emotion-enhance",
+        headers={"Authorization": api_key, "Content-Type": "application/json"},
+        json={"text": text},
+        timeout=timeout,
+    )
+    if resp.status_code != 200:
+        raise RuntimeError(
+            f"/emotion-enhance failed: status={resp.status_code}, body={resp.text}"
+        )
+    enhanced = resp.json().get("data", {}).get("emotion_enhance")
+    if not enhanced:
+        raise RuntimeError(f"/emotion-enhance returned no data: {resp.text}")
+    return enhanced
+
+
+def _bool_form(v: Any) -> str:
+    return "true" if bool(v) else "false"
+
+
+def _resolve_reference_audio(ref: str, timeout: int) -> Tuple[Path, Optional[Path]]:
+    """Resolve reference_audio to a path. If ref is a URL, download to temp file.
+    Returns (path_to_use, temp_path_to_cleanup_or_None)."""
+    if ref.startswith("http://") or ref.startswith("https://"):
+        import requests
+        tmp = tempfile.NamedTemporaryFile(suffix=".wav", delete=False)
+        tmp.close()
+        r = requests.get(ref, timeout=timeout)
+        r.raise_for_status()
+        Path(tmp.name).write_bytes(r.content)
+        return Path(tmp.name), Path(tmp.name)
+    p = Path(ref)
+    if not p.exists():
+        raise FileNotFoundError(f"reference_audio not found: {ref}")
+    return p, None
+
+
+def _noiz_tts(
+    base_url: str,
+    api_key: str,
+    cue: Cue,
+    cfg: Dict[str, Any],
+    output_format: str,
+    timeout: int,
+    out_path: Path,
+) -> float:
+    import requests
+
+    url = f"{base_url.rstrip('/')}/text-to-speech"
+    payload: Dict[str, str] = {
+        "text": cue.text,
+        "duration": f"{cue.duration_ms / 1000.0:.3f}",
+        "output_format": output_format,
+    }
+    for field in ("voice_id", "quality_preset", "speed", "target_lang"):
+        if field in cfg and cfg[field] is not None:
+            payload[field] = str(cfg[field])
+    if "similarity_enh" in cfg:
+        payload["similarity_enh"] = _bool_form(cfg["similarity_enh"])
+    if "save_voice" in cfg:
+        payload["save_voice"] = _bool_form(cfg["save_voice"])
+    if "emo" in cfg and cfg["emo"] is not None:
+        emo = cfg["emo"]
+        payload["emo"] = emo if isinstance(emo, str) else json.dumps(emo)
+
+    files = None
+    ref_cleanup: Optional[Path] = None
+    ref = cfg.get("reference_audio")
+    if ref:
+        ref_path, ref_cleanup = _resolve_reference_audio(ref, timeout)
+        files = {
+            "file": (
+                ref_path.name,
+                ref_path.open("rb"),
+                "application/octet-stream",
+            )
+        }
+    elif not cfg.get("voice_id"):
+        raise ValueError(
+            f"Cue {cue.index}: either voice_id or reference_audio required."
+        )
+
+    try:
+        resp = requests.post(
+            url, headers={"Authorization": api_key},
+            data=payload, files=files, timeout=timeout,
+        )
+    finally:
+        if files and files["file"][1]:
+            files["file"][1].close()
+        if ref_cleanup is not None:
+            ref_cleanup.unlink(missing_ok=True)
+
+    if resp.status_code != 200:
+        raise RuntimeError(
+            f"/text-to-speech cue {cue.index}: "
+            f"status={resp.status_code}, body={resp.text}"
+        )
+    out_path.write_bytes(resp.content)
+    dur_h = resp.headers.get("X-Audio-Duration")
+    return float(dur_h) if dur_h else -1.0
+
+
+# ── Kokoro backend ───────────────────────────────────────────────────
+
+
+def _ensure_kokoro() -> None:
+    if not shutil.which("kokoro-tts"):
+        raise RuntimeError("kokoro-tts CLI not found.")
+
+
+def _kokoro_tts(
+    cue: Cue,
+    cfg: Dict[str, Any],
+    output_format: str,
+    out_path: Path,
+) -> float:
+    with tempfile.NamedTemporaryFile(
+        mode="w", suffix=".txt", delete=False, encoding="utf-8"
+    ) as tmp:
+        tmp.write(cue.text)
+        tmp_path = tmp.name
+
+    try:
+        cmd = ["kokoro-tts", tmp_path, str(out_path)]
+        voice = cfg.get("voice")
+        if voice:
+            cmd += ["--voice", str(voice)]
+        lang = cfg.get("lang")
+        if lang:
+            cmd += ["--lang", str(lang)]
+        speed = cfg.get("speed")
+        if speed is not None:
+            cmd += ["--speed", str(speed)]
+        cmd += ["--format", output_format]
+
+        proc = subprocess.run(cmd, capture_output=True, text=True)
+        if proc.returncode != 0:
+            raise RuntimeError(
+                f"kokoro-tts failed for cue {cue.index}: {proc.stderr}"
+            )
+    finally:
+        Path(tmp_path).unlink(missing_ok=True)
+
+    if out_path.exists():
+        return probe_duration_ms(out_path) / 1000.0
+    raise RuntimeError(f"kokoro-tts produced no output for cue {cue.index}")
+
+
+# ── main ─────────────────────────────────────────────────────────────
+
+
+def main() -> int:
+    ap = argparse.ArgumentParser(
+        description="Render timeline-accurate speech from SRT."
+    )
+    ap.add_argument("--srt", required=True, help="Input SRT file")
+    ap.add_argument("--voice-map", required=True, help="Voice-map JSON")
+    ap.add_argument(
+        "--backend", choices=["kokoro", "noiz"], default="kokoro",
+        help="TTS backend (default: kokoro)",
+    )
+    ap.add_argument("--api-key", help="API key (required for noiz backend)")
+    ap.add_argument("--output", required=True, help="Output audio file")
+    ap.add_argument("--base-url", default="https://noiz.ai/v1")
+    ap.add_argument("--work-dir", default=".tmp/tts")
+    ap.add_argument("--auto-emotion", action="store_true",
+                     help="Noiz backend only: call /emotion-enhance before TTS")
+    ap.add_argument("--ref-audio-track", help="Original audio track to dynamically slice as reference audio per segment")
+    ap.add_argument("--output-format", choices=["wav", "mp3"], default="wav")
+    ap.add_argument("--timeout-sec", type=int, default=120)
+    args = ap.parse_args()
+
+    if args.backend == "noiz" and not args.api_key:
+        print("Error: --api-key is required for noiz backend.", file=sys.stderr)
+        return 1
+    if args.api_key:
+        args.api_key = normalize_api_key_base64(args.api_key)
+
+    try:
+        ensure_ffmpeg()
+        if args.backend == "kokoro":
+            _ensure_kokoro()
+
+        work = Path(args.work_dir)
+        work.mkdir(parents=True, exist_ok=True)
+
+        cues = parse_srt(Path(args.srt))
+        voice_map = json.loads(Path(args.voice_map).read_text(encoding="utf-8"))
+
+        delayed: List[Path] = []
+        report: List[Dict[str, Any]] = []
+
+        for cue in cues:
+            cfg = resolve_segment_cfg(cue.index, voice_map)
+            
+            if args.ref_audio_track and not cfg.get("voice_id") and not cfg.get("reference_audio"):
+                ref_slice_path = work / f"seg_{cue.index:04d}_ref.wav"
+                if not ref_slice_path.exists():
+                    _run_ff([
+                        "ffmpeg", "-y",
+                        "-ss", f"{cue.start_ms / 1000.0:.3f}",
+                        "-i", str(args.ref_audio_track),
+                        "-t", f"{cue.duration_ms / 1000.0:.3f}",
+                        "-acodec", "pcm_s16le", "-ar", "16000", "-ac", "1",
+                        str(ref_slice_path)
+                    ])
+                cfg["reference_audio"] = str(ref_slice_path)
+            
+            text = cue.text
+
+            if args.backend == "noiz" and args.auto_emotion:
+                text = _noiz_emotion_enhance(
+                    args.base_url, args.api_key, cue.text, args.timeout_sec
+                )
+
+            synth_cue = Cue(cue.index, cue.start_ms, cue.end_ms, text)
+            raw = work / f"seg_{cue.index:04d}_raw.{args.output_format}"
+            norm = work / f"seg_{cue.index:04d}_norm.wav"
+            dly = work / f"seg_{cue.index:04d}_delay.wav"
+
+            if args.backend == "noiz":
+                api_dur = _noiz_tts(
+                    args.base_url, args.api_key, synth_cue,
+                    cfg, args.output_format, args.timeout_sec, raw,
+                )
+                normalize_duration_pad_trim(raw, norm, cue.duration_ms)
+            else:
+                api_dur = _kokoro_tts(synth_cue, cfg, args.output_format, raw)
+                normalize_duration_atempo(raw, norm, cue.duration_ms)
+
+            delay_segment(norm, dly, cue.start_ms)
+            delayed.append(dly)
+
+            seg_report: Dict[str, Any] = {
+                "index": cue.index,
+                "start_ms": cue.start_ms,
+                "end_ms": cue.end_ms,
+                "duration_ms": cue.duration_ms,
+                "raw_duration_sec": api_dur,
+                "backend": args.backend,
+            }
+            if args.backend == "noiz":
+                seg_report["voice_id"] = cfg.get("voice_id")
+                seg_report["reference_audio"] = cfg.get("reference_audio")
+                seg_report["emo"] = cfg.get("emo")
+            else:
+                seg_report["voice"] = cfg.get("voice")
+                seg_report["lang"] = cfg.get("lang")
+            report.append(seg_report)
+
+        timeline_wav = work / "timeline.wav"
+        total_ms = max(c.end_ms for c in cues)
+        mix_all(delayed, timeline_wav, total_ms)
+
+        out = Path(args.output)
+        if out.suffix.lower() != ".wav":
+            _run_ff(["ffmpeg", "-y", "-i", str(timeline_wav), str(out)])
+        else:
+            out.parent.mkdir(parents=True, exist_ok=True)
+            out.write_bytes(timeline_wav.read_bytes())
+
+        report_path = work / "render_report.json"
+        report_path.write_text(
+            json.dumps({
+                "srt": args.srt,
+                "output": args.output,
+                "backend": args.backend,
+                "total_ms": total_ms,
+                "segments": report,
+            }, ensure_ascii=False, indent=2),
+            encoding="utf-8",
+        )
+        print(f"Done. Output: {out}")
+        print(f"Report: {report_path}")
+        return 0
+    except Exception as exc:
+        print(f"Error: {exc}", file=sys.stderr)
+        return 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())

skills/speak/scripts/text_to_srt.py

@@ -0,0 +1,115 @@
+#!/usr/bin/env python3
+"""Convert plain text to SRT with auto-estimated timings.
+
+Splits text into sentences, estimates duration per sentence based on
+character-per-second rate, and writes a valid SRT file.
+"""
+import argparse
+import re
+import sys
+from pathlib import Path
+from typing import List, Tuple
+
+
+SENTENCE_SPLIT_RE = re.compile(
+    r'(?<=[。！？.!?\n])\s*'
+)
+
+
+def split_sentences(text: str) -> List[str]:
+    raw = SENTENCE_SPLIT_RE.split(text.strip())
+    sentences = [s.strip() for s in raw if s.strip()]
+    return sentences
+
+
+def estimate_timings(
+    sentences: List[str],
+    chars_per_second: float,
+    gap_ms: int,
+    start_offset_ms: int = 0,
+) -> List[Tuple[int, int, int, str]]:
+    """Return list of (index, start_ms, end_ms, text)."""
+    result = []
+    cursor_ms = start_offset_ms
+    for i, sentence in enumerate(sentences, start=1):
+        char_count = len(sentence)
+        duration_ms = max(500, int(char_count / chars_per_second * 1000))
+        start_ms = cursor_ms
+        end_ms = start_ms + duration_ms
+        result.append((i, start_ms, end_ms, sentence))
+        cursor_ms = end_ms + gap_ms
+    return result
+
+
+def ms_to_srt_time(ms: int) -> str:
+    total_sec, millis = divmod(ms, 1000)
+    total_min, sec = divmod(total_sec, 60)
+    hour, minute = divmod(total_min, 60)
+    return f"{hour:02d}:{minute:02d}:{sec:02d},{millis:03d}"
+
+
+def write_srt(entries: List[Tuple[int, int, int, str]], path: Path) -> None:
+    lines = []
+    for idx, start_ms, end_ms, text in entries:
+        lines.append(str(idx))
+        lines.append(f"{ms_to_srt_time(start_ms)} --> {ms_to_srt_time(end_ms)}")
+        lines.append(text)
+        lines.append("")
+    path.parent.mkdir(parents=True, exist_ok=True)
+    path.write_text("\n".join(lines), encoding="utf-8")
+
+
+def main() -> int:
+    parser = argparse.ArgumentParser(
+        description="Convert text file to SRT with auto-estimated timings."
+    )
+    parser.add_argument("--input", required=True, help="Input text file path")
+    parser.add_argument("--output", required=True, help="Output SRT file path")
+    parser.add_argument(
+        "--chars-per-second",
+        type=float,
+        default=4.0,
+        help="Reading speed in characters per second (default: 4.0, good for Chinese; "
+        "use ~15 for English)",
+    )
+    parser.add_argument(
+        "--gap-ms",
+        type=int,
+        default=300,
+        help="Gap between segments in milliseconds (default: 300)",
+    )
+    parser.add_argument(
+        "--start-offset-ms",
+        type=int,
+        default=0,
+        help="Timeline start offset in milliseconds (default: 0)",
+    )
+    args = parser.parse_args()
+
+    try:
+        text = Path(args.input).read_text(encoding="utf-8").strip()
+        if not text:
+            raise ValueError("Input text is empty.")
+
+        sentences = split_sentences(text)
+        if not sentences:
+            raise ValueError("No sentences found after splitting.")
+
+        entries = estimate_timings(
+            sentences,
+            chars_per_second=args.chars_per_second,
+            gap_ms=args.gap_ms,
+            start_offset_ms=args.start_offset_ms,
+        )
+        write_srt(entries, Path(args.output))
+        print(f"Done. {len(entries)} segments written to {args.output}")
+        total_ms = entries[-1][2] if entries else 0
+        print(f"Total duration: {ms_to_srt_time(total_ms)}")
+        return 0
+    except Exception as exc:
+        print(f"Error: {exc}", file=sys.stderr)
+        return 1
+
+
+if __name__ == "__main__":
+    raise SystemExit(main())