File size: 12,837 Bytes
9fc9873 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 | #!/usr/bin/env python3
"""Generate and audit the small HFD emergency-stop/safe-hold TTS corpus.
Generation is shardable by voice so four local GPUs can run independently.
The final manifest is written only by ``--mode finalize`` after every expected
file passes the exact PCM16/16 kHz contract consumed by HFD training.
"""
from __future__ import annotations
import argparse
import hashlib
import json
import math
import os
import wave
from pathlib import Path
from typing import Any
import numpy as np
FORMAT = "hfd_control_command_tts_v1"
TARGET_RATE = 16_000
DEFAULT_MODEL = Path(
"/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/team_workspace/public_ckpt/Qwen/"
"Qwen3-TTS-12Hz-1.7B-VoiceDesign"
)
DEFAULT_OUTPUT = Path(
"/cpfs01/thscc/sharestorage/AIGC-06-01/Hithink_VideoGen_Group/user_workspace/yetinghong/videogen/hfd_control_commands_v1"
)
# Each phrase ID represents one bilingual meaning. Every meaning is rendered
# by female/male × speaker 01/02 × Chinese/English = exactly eight WAVs.
PHRASES: tuple[dict[str, str], ...] = (
{
"intent": "emergency-stop",
"phrase_id": "stop_01",
"zh": "停。",
"en": "Stop.",
},
{
"intent": "emergency-stop",
"phrase_id": "stop_02",
"zh": "马上停下来。",
"en": "Stop now.",
},
{
"intent": "emergency-stop",
"phrase_id": "stop_03",
"zh": "先停一下。",
"en": "Pause for a moment.",
},
{
"intent": "emergency-stop",
"phrase_id": "stop_04",
"zh": "停止动作。",
"en": "Stop the movement.",
},
{
"intent": "emergency-stop",
"phrase_id": "stop_05",
"zh": "立刻停止。",
"en": "Stop immediately.",
},
{
"intent": "emergency-stop",
"phrase_id": "stop_06",
"zh": "不要再继续了。",
"en": "Do not continue.",
},
{
"intent": "safe-hold",
"phrase_id": "hold_01",
"zh": "别动。",
"en": "Don't move.",
},
{
"intent": "safe-hold",
"phrase_id": "hold_02",
"zh": "不要动。",
"en": "Stay still.",
},
{
"intent": "safe-hold",
"phrase_id": "hold_03",
"zh": "不对,先别动。",
"en": "That's not right. Hold still.",
},
{
"intent": "safe-hold",
"phrase_id": "hold_04",
"zh": "保持现在的姿势。",
"en": "Hold your current pose.",
},
{
"intent": "safe-hold",
"phrase_id": "hold_05",
"zh": "先保持不动。",
"en": "Keep still for now.",
},
{
"intent": "safe-hold",
"phrase_id": "hold_06",
"zh": "站稳,别动。",
"en": "Stay balanced and don't move.",
},
)
def _voice_specs() -> tuple[dict[str, str], ...]:
return (
{
"name": "female_01_en",
"language": "en",
"qwen_language": "English",
"instruction": (
"Use a natural everyday adult female voice. Speak the short "
"urgent command clearly and calmly, without theatrical acting."
),
},
{
"name": "female_02_en",
"language": "en",
"qwen_language": "English",
"instruction": (
"Use a natural young adult female voice. Speak clearly in a "
"concise, neutral, everyday tone."
),
},
{
"name": "male_01_en",
"language": "en",
"qwen_language": "English",
"instruction": (
"Use a natural everyday adult male voice. Speak the short "
"command clearly and firmly, but do not shout."
),
},
{
"name": "male_02_en",
"language": "en",
"qwen_language": "English",
"instruction": (
"Use a natural young adult male voice. Speak concisely in a "
"relaxed, neutral conversational tone."
),
},
{
"name": "female_01_zh",
"language": "zh",
"qwen_language": "Chinese",
"instruction": "使用自然日常的成年女性声音,清晰简短地说出指令,语气坚定但不要喊叫。",
},
{
"name": "female_02_zh",
"language": "zh",
"qwen_language": "Chinese",
"instruction": "使用自然日常的年轻女性声音,语速适中,语气中性简洁,不要戏剧化。",
},
{
"name": "male_01_zh",
"language": "zh",
"qwen_language": "Chinese",
"instruction": "使用自然日常的成年男性声音,清晰简短地说出指令,坚定但不要大喊。",
},
{
"name": "male_02_zh",
"language": "zh",
"qwen_language": "Chinese",
"instruction": "使用自然日常的年轻男性声音,语气放松中性,简洁自然地说出指令。",
},
)
def _audio_path(root: Path, phrase: dict[str, str], voice: dict[str, str]) -> Path:
return root / "audio" / phrase["intent"] / phrase["phrase_id"] / f"{voice['name']}.wav"
def _sha256(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(1024 * 1024), b""):
digest.update(chunk)
return digest.hexdigest()
def _atomic_jsonl(path: Path, rows: list[dict[str, Any]]) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
temporary = path.with_suffix(path.suffix + f".tmp.{os.getpid()}")
with temporary.open("w", encoding="utf-8") as handle:
for row in rows:
handle.write(json.dumps(row, ensure_ascii=False, sort_keys=True) + "\n")
handle.flush()
os.fsync(handle.fileno())
temporary.replace(path)
def _write_pcm16(path: Path, waveform: np.ndarray) -> None:
values = np.asarray(waveform, dtype=np.float32).reshape(-1)
if values.size <= 0 or not np.isfinite(values).all():
raise RuntimeError(f"TTS produced invalid waveform for {path}")
peak = float(np.max(np.abs(values), initial=0.0))
if peak > 1.0:
values = values / np.float32(peak)
pcm = np.rint(np.clip(values, -1.0, 1.0) * 32767.0).astype("<i2")
path.parent.mkdir(parents=True, exist_ok=True)
temporary = path.with_suffix(f".tmp.{os.getpid()}.wav")
with wave.open(str(temporary), "wb") as handle:
handle.setnchannels(1)
handle.setsampwidth(2)
handle.setframerate(TARGET_RATE)
handle.writeframes(pcm.tobytes())
temporary.replace(path)
def _valid_pcm16(path: Path) -> bool:
if not path.is_file():
return False
try:
with wave.open(str(path), "rb") as handle:
return (
handle.getnchannels() == 1
and handle.getsampwidth() == 2
and handle.getframerate() == TARGET_RATE
and handle.getnframes() > 0
and handle.getcomptype() == "NONE"
)
except (EOFError, wave.Error):
return False
def generate(args: argparse.Namespace) -> None:
import torch
from scipy.signal import resample_poly
from qwen_tts import Qwen3TTSModel
if not 0 <= args.shard_index < args.shard_count:
raise ValueError(
f"Invalid shard {args.shard_index}/{args.shard_count}"
)
voices = tuple(
voice
for index, voice in enumerate(_voice_specs())
if index % args.shard_count == args.shard_index
)
if not voices:
print(f"[ready] shard={args.shard_index} has no assigned voices")
return
torch.backends.cuda.enable_cudnn_sdp(False)
torch.backends.cuda.enable_flash_sdp(True)
torch.backends.cuda.enable_mem_efficient_sdp(True)
torch.backends.cuda.enable_math_sdp(True)
torch.backends.cuda.matmul.allow_tf32 = True
model = Qwen3TTSModel.from_pretrained(
str(args.model_path),
device_map=args.device,
dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
)
for voice in voices:
pending = [
phrase
for phrase in PHRASES
if not _valid_pcm16(_audio_path(args.output_root, phrase, voice))
]
for start in range(0, len(pending), max(args.batch_size, 1)):
batch = pending[start : start + max(args.batch_size, 1)]
texts = [phrase[voice["language"]] for phrase in batch]
waveforms, source_rate = model.generate_voice_design(
text=texts,
language=[voice["qwen_language"]] * len(batch),
instruct=[voice["instruction"]] * len(batch),
)
if len(waveforms) != len(batch):
raise RuntimeError("Qwen TTS returned the wrong batch size")
for phrase, waveform in zip(batch, waveforms):
values = np.asarray(waveform, dtype=np.float32).reshape(-1)
if int(source_rate) != TARGET_RATE:
divisor = math.gcd(int(source_rate), TARGET_RATE)
values = resample_poly(
values,
TARGET_RATE // divisor,
int(source_rate) // divisor,
).astype(np.float32, copy=False)
_write_pcm16(_audio_path(args.output_root, phrase, voice), values)
print(
f"[tts] shard={args.shard_index}/{args.shard_count} "
f"voice={voice['name']} done={min(start + len(batch), len(pending))}/"
f"{len(pending)}",
flush=True,
)
del model
torch.cuda.empty_cache()
def finalize(args: argparse.Namespace) -> None:
rows: list[dict[str, Any]] = []
for phrase in PHRASES:
for voice in _voice_specs():
path = _audio_path(args.output_root, phrase, voice).resolve()
if not _valid_pcm16(path):
raise RuntimeError(f"Missing/invalid expected control WAV: {path}")
with wave.open(str(path), "rb") as handle:
samples = int(handle.getnframes())
raw = handle.readframes(samples)
values = np.frombuffer(raw, dtype="<i2").astype(np.float32) / 32768.0
peak = float(np.max(np.abs(values), initial=0.0))
speech_scale = (
min(1.0, (10.0 ** (-3.0 / 20.0)) / peak)
if peak > 0.0
else 1.0
)
rows.append(
{
"format": FORMAT,
"intent": phrase["intent"],
"phrase_id": phrase["phrase_id"],
"voice_name": voice["name"],
"language": voice["language"],
"text": phrase[voice["language"]],
"audio_path": str(path),
"sample_rate": TARGET_RATE,
"speech_samples_16k": samples,
"duration_seconds": samples / float(TARGET_RATE),
"peak": peak,
"speech_scale": speech_scale,
"sha256": _sha256(path),
}
)
expected = len(PHRASES) * len(_voice_specs())
if len(rows) != expected:
raise AssertionError(f"Manifest row count drifted: {len(rows)} != {expected}")
_atomic_jsonl(args.output_root / "manifest.jsonl", rows)
durations = [float(row["duration_seconds"]) for row in rows]
print(
f"[ready] rows={len(rows)} meanings={len(PHRASES)} voices=8 "
f"duration={min(durations):.2f}-{max(durations):.2f}s "
f"manifest={args.output_root / 'manifest.jsonl'}",
flush=True,
)
def parse_args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--mode", choices=("generate", "finalize"), required=True)
parser.add_argument("--output-root", type=Path, default=DEFAULT_OUTPUT)
parser.add_argument("--model-path", type=Path, default=DEFAULT_MODEL)
parser.add_argument("--device", default="cuda:0")
parser.add_argument("--batch-size", type=int, default=6)
parser.add_argument("--shard-index", type=int, default=0)
parser.add_argument("--shard-count", type=int, default=1)
return parser.parse_args()
def main() -> None:
args = parse_args()
args.output_root = args.output_root.expanduser().resolve()
args.model_path = args.model_path.expanduser().resolve()
if args.mode == "generate":
generate(args)
else:
finalize(args)
if __name__ == "__main__":
main()
|