Spaces:
Running on Zero
Running on Zero
File size: 16,871 Bytes
0422215 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 | """
Step 6: Audio sync β match synthesised segment durations to original timestamps.
For each segment:
- Too long β speed up using ffmpeg atempo filter
- Too short β pad with silence at the end
Then stitch all segments into a single final audio track.
"""
import array
import math
import os
import subprocess
import wave
from pathlib import Path
def _get_wav_duration(wav_path: str) -> float:
with wave.open(wav_path, 'r') as f:
frames = f.getnframes()
rate = f.getframerate()
return frames / float(rate)
def _speedup_audio(input_path: str, output_path: str, factor: float) -> None:
"""Speed up/slow down audio by factor using ffmpeg atempo (supports 0.5β100x via chaining)."""
# atempo supports 0.5 to 2.0, chain filters for larger factors
filters = []
remaining = factor
while remaining > 2.0:
filters.append("atempo=2.0")
remaining /= 2.0
while remaining < 0.5:
filters.append("atempo=0.5")
remaining /= 0.5
filters.append(f"atempo={remaining:.4f}")
filter_str = ",".join(filters)
cmd = [
"ffmpeg", "-y", "-i", input_path,
"-filter:a", filter_str,
output_path,
]
result = subprocess.run(cmd, capture_output=True, text=True)
if result.returncode != 0:
raise RuntimeError(f"ffmpeg atempo failed:\n{result.stderr}")
def _pad_silence(input_path: str, output_path: str, target_duration: float) -> None:
"""Pad audio with silence to reach target_duration seconds."""
current = _get_wav_duration(input_path)
pad_seconds = max(0, target_duration - current)
cmd = [
"ffmpeg", "-y", "-i", input_path,
"-af", f"apad=pad_dur={pad_seconds:.4f}",
"-t", str(target_duration),
output_path,
]
result = subprocess.run(cmd, capture_output=True, text=True)
if result.returncode != 0:
raise RuntimeError(f"ffmpeg apad failed:\n{result.stderr}")
def _trim_audio(input_path: str, output_path: str, duration: float) -> None:
"""Trim audio to exactly duration seconds."""
tmp = output_path + ".trim.wav"
cmd = ["ffmpeg", "-y", "-i", input_path, "-t", str(duration), tmp]
result = subprocess.run(cmd, capture_output=True, text=True)
if result.returncode != 0:
raise RuntimeError(f"ffmpeg trim failed:\n{result.stderr}")
os.replace(tmp, output_path)
def _detect_pauses(words: list[dict], min_pause: float = 0.15) -> list[dict]:
"""Find gaps between consecutive words that exceed min_pause seconds.
Returns list of {after_word_idx, position, duration} sorted by position.
"""
pauses = []
for i in range(len(words) - 1):
gap = words[i + 1]["start"] - words[i]["end"]
if gap >= min_pause:
pauses.append({
"after_word_idx": i,
"position": words[i]["end"],
"duration": gap,
})
return pauses
def _find_tts_silences(wav_path: str, threshold_db: float = -35.0,
min_dur: float = 0.08) -> list[dict]:
"""Find silence regions in a TTS WAV using RMS energy.
Returns list of {start, end, duration} for each detected silence region.
"""
with wave.open(wav_path, "r") as f:
n_frames = f.getnframes()
sample_rate = f.getframerate()
raw = f.readframes(n_frames)
# Convert raw bytes to 16-bit signed samples
samples = array.array("h", raw)
window_size = int(0.02 * sample_rate) # 20 ms windows
hop = window_size // 2
threshold_linear = 10 ** (threshold_db / 20.0) * 32768 # dBFS to linear amplitude
silences: list[dict] = []
in_silence = False
silence_start = 0.0
for pos in range(0, len(samples) - window_size, hop):
chunk = samples[pos:pos + window_size]
rms = math.sqrt(sum(s * s for s in chunk) / window_size)
t = pos / sample_rate
if rms < threshold_linear:
if not in_silence:
in_silence = True
silence_start = t
else:
if in_silence:
dur = t - silence_start
if dur >= min_dur:
silences.append({"start": silence_start, "end": t, "duration": dur})
in_silence = False
# Close trailing silence
if in_silence:
t_end = len(samples) / sample_rate
dur = t_end - silence_start
if dur >= min_dur:
silences.append({"start": silence_start, "end": t_end, "duration": dur})
return silences
def _read_wav_samples(wav_path: str) -> tuple[array.array, int]:
"""Read a mono 16-bit WAV and return (samples, sample_rate)."""
with wave.open(wav_path, "r") as f:
sr = f.getframerate()
raw = f.readframes(f.getnframes())
return array.array("h", raw), sr
def _write_wav_samples(samples: array.array, sample_rate: int, output_path: str) -> None:
"""Write 16-bit mono samples to a WAV file."""
with wave.open(output_path, "w") as f:
f.setnchannels(1)
f.setsampwidth(2)
f.setframerate(sample_rate)
f.writeframes(samples.tobytes())
def _pause_aware_sync(tts_path: str, synced_path: str, target_duration: float,
words: list[dict], max_speed: float,
max_overflow: float = 0.0) -> None:
"""Sync TTS audio using pause-aware strategy: compress silences first, then atempo.
When TTS is too long: shrink detected silence regions before speeding up speech.
When TTS is too short: distribute extra padding at natural pause points.
`max_overflow`: extra seconds the synced output may exceed target_duration without
trimming. The caller borrows this budget from the inter-segment silence that follows,
so we never silently drop trailing words just to hit `target_duration` exactly.
"""
tts_duration = _get_wav_duration(tts_path)
original_pauses = _detect_pauses(words)
tts_silences = _find_tts_silences(tts_path)
total_tts_silence = sum(s["duration"] for s in tts_silences)
hard_cap = target_duration + max_overflow
overshoot_vs_cap = tts_duration - hard_cap
if tts_duration > target_duration * 1.02:
if tts_silences and total_tts_silence > 0:
if overshoot_vs_cap <= 0:
# Already within hard_cap once we factor in the borrow budget β keep TTS as-is.
import shutil
shutil.copy(tts_path, synced_path)
print(f"[s5] pause-aware: within +{max_overflow:.2f}s borrow, no compression")
else:
removable = min(total_tts_silence * 0.9, overshoot_vs_cap)
if removable >= overshoot_vs_cap:
compression_ratio = 1.0 - (removable / total_tts_silence)
_compress_silences(tts_path, synced_path, tts_silences, compression_ratio)
print(f"[s5] pause-aware: compressed silences (ratio {compression_ratio:.2f}, +{max_overflow:.2f}s borrow)")
else:
_compress_silences(tts_path, synced_path, tts_silences, 0.1) # keep 10%
remaining_dur = _get_wav_duration(synced_path)
speed_factor = remaining_dur / hard_cap if hard_cap > 0 else max_speed
if speed_factor > max_speed:
print(f"[s5] pause-aware: WARNING speed x{speed_factor:.2f} exceeds max, capping at x{max_speed} (will overflow next gap)")
speed_factor = max_speed
print(f"[s5] pause-aware: compressed silences + speedup x{speed_factor:.2f} (+{max_overflow:.2f}s borrow)")
tmp = synced_path + ".tmp.wav"
_speedup_audio(synced_path, tmp, speed_factor)
os.replace(tmp, synced_path)
else:
# No silences detected β uniform speedup, but use hard_cap as the target.
speed_factor = tts_duration / hard_cap if hard_cap > 0 else max_speed
if speed_factor > max_speed:
print(f"[s5] pause-aware: WARNING speed x{speed_factor:.2f} exceeds max, capping at x{max_speed} (will overflow next gap)")
speed_factor = max_speed
print(f"[s5] pause-aware: uniform speedup x{speed_factor:.2f} (no silences, +{max_overflow:.2f}s borrow)")
_speedup_audio(tts_path, synced_path, speed_factor)
elif tts_duration < target_duration * 0.98:
shortfall = target_duration - tts_duration
if tts_silences and original_pauses:
# Distribute padding at detected silence positions
_distribute_padding(tts_path, synced_path, tts_silences, shortfall, target_duration)
print(f"[s5] pause-aware: distributed {shortfall:.2f}s padding across {len(tts_silences)} pause points")
else:
# No pause points β pad at end
_pad_silence(tts_path, synced_path, target_duration)
print(f"[s5] pause-aware: padded {shortfall:.2f}s at end (no pause points)")
else:
import shutil
shutil.copy(tts_path, synced_path)
def _compress_silences(input_path: str, output_path: str,
silences: list[dict], keep_ratio: float) -> None:
"""Rewrite WAV with silence regions compressed to keep_ratio of their original duration."""
samples, sr = _read_wav_samples(input_path)
out = array.array("h")
prev_end_sample = 0
for sil in silences:
sil_start = int(sil["start"] * sr)
sil_end = int(sil["end"] * sr)
# Copy speech before this silence
out.extend(samples[prev_end_sample:sil_start])
# Keep only keep_ratio of the silence
kept_samples = int((sil_end - sil_start) * keep_ratio)
if kept_samples > 0:
out.extend(samples[sil_start:sil_start + kept_samples])
prev_end_sample = sil_end
# Copy remaining speech after last silence
out.extend(samples[prev_end_sample:])
_write_wav_samples(out, sr, output_path)
def _distribute_padding(input_path: str, output_path: str,
tts_silences: list[dict], shortfall: float,
target_duration: float) -> None:
"""Insert extra silence distributed across detected pause points."""
samples, sr = _read_wav_samples(input_path)
n_points = len(tts_silences)
pad_per_point = shortfall / n_points
out = array.array("h")
prev_end_sample = 0
for sil in tts_silences:
sil_end = int(sil["end"] * sr)
# Copy everything up to end of this silence region
out.extend(samples[prev_end_sample:sil_end])
# Insert extra silence
extra_samples = int(pad_per_point * sr)
out.extend(array.array("h", [0] * extra_samples))
prev_end_sample = sil_end
# Copy remaining audio
out.extend(samples[prev_end_sample:])
_write_wav_samples(out, sr, output_path)
# Trim to exact target if slightly over due to rounding
actual = len(out) / sr
if actual > target_duration * 1.02:
_trim_audio(output_path, output_path, target_duration)
def _generate_silence(output_path: str, duration: float, sample_rate: int = 16000) -> None:
"""Generate a silent WAV file of given duration."""
num_samples = int(duration * sample_rate)
with wave.open(output_path, "w") as f:
f.setnchannels(1)
f.setsampwidth(2) # 16-bit
f.setframerate(sample_rate)
f.writeframes(b"\x00\x00" * num_samples)
def sync_and_stitch(
segments: list[dict],
output_path: str = "tmp/audio/final_audio.wav",
synced_dir: str = "tmp/audio/tts_synced",
max_speed: float = 1.8,
) -> str:
"""
Sync each TTS segment to its original timestamp window and stitch into a single WAV.
Args:
segments: List of dicts with {start, end, tts_path}.
output_path: Where to write the final stitched audio.
synced_dir: Temp directory for per-segment synced WAVs.
max_speed: Maximum allowed speedup factor (default 1.8x to preserve naturalness).
Returns:
Path to the final stitched audio WAV.
"""
Path(synced_dir).mkdir(parents=True, exist_ok=True)
Path(output_path).parent.mkdir(parents=True, exist_ok=True)
# Detect TTS sample rate from the first segment
with wave.open(segments[0]["tts_path"], 'r') as f:
tts_sample_rate = f.getframerate()
print(f"[s5] TTS sample rate: {tts_sample_rate} Hz")
concat_list_path = "tmp/concat_list.txt"
concat_entries = []
# Track the real wall-clock playback cursor. When a segment overflows its
# original window, the cursor moves past the segment's nominal end, and the
# next inter-segment silence shrinks accordingly β overflow is absorbed by
# the following gap instead of being trimmed off the end of the audio.
playback_cursor = 0.0
for i, seg in enumerate(segments):
start = seg["start"]
end = seg["end"]
target_duration = end - start
tts_path = seg["tts_path"]
# Fill gap before this segment with silence β but only as much as the
# cursor is actually behind. If a prior segment overflowed past `start`,
# `gap` goes negative and we skip the silence (and start slightly late).
gap = start - playback_cursor
if gap > 0.01:
sil_path = os.path.join(synced_dir, f"silence_{i:04d}.wav")
_generate_silence(sil_path, gap, sample_rate=tts_sample_rate)
concat_entries.append(sil_path)
playback_cursor += gap
elif gap < -0.05:
print(f"[s5] Seg {i}: running {-gap:.2f}s behind original timeline (prior overflow absorbed)")
# Borrow budget: how much we may overflow `target_duration` without
# trimming. We can use the silence between this segment's `end` and the
# next segment's `start`. Last segment has no follower β 0.
if i + 1 < len(segments):
allowed_overflow = max(segments[i + 1]["start"] - end, 0.0)
else:
allowed_overflow = 0.0
tts_duration = _get_wav_duration(tts_path)
synced_path = os.path.join(synced_dir, f"synced_{i:04d}.wav")
hard_cap = target_duration + allowed_overflow
words = seg.get("words")
if words and len(words) > 1:
print(f"[s5] Seg {i}: pause-aware sync ({tts_duration:.2f}s -> {target_duration:.2f}s, +{allowed_overflow:.2f}s borrow)")
_pause_aware_sync(tts_path, synced_path, target_duration, words, max_speed,
max_overflow=allowed_overflow)
elif tts_duration > target_duration * 1.02:
# Speed up only as far as needed to land within hard_cap; if the
# required factor exceeds max_speed, cap it and let it overflow β
# the next gap will shrink to absorb it. Never trim.
if tts_duration <= hard_cap:
speed_factor = 1.0
else:
speed_factor = tts_duration / hard_cap if hard_cap > 0 else max_speed
if speed_factor > max_speed:
print(f"[s5] Seg {i}: WARNING speed x{speed_factor:.2f} exceeds max, capping at x{max_speed} (will overflow next gap)")
speed_factor = max_speed
if speed_factor > 1.001:
print(f"[s5] Seg {i}: speeding up x{speed_factor:.2f} (+{allowed_overflow:.2f}s borrow)")
_speedup_audio(tts_path, synced_path, speed_factor)
else:
import shutil
shutil.copy(tts_path, synced_path)
print(f"[s5] Seg {i}: within +{allowed_overflow:.2f}s borrow, no speedup")
elif tts_duration < target_duration * 0.98:
print(f"[s5] Seg {i}: padding {target_duration - tts_duration:.2f}s silence")
_pad_silence(tts_path, synced_path, target_duration)
else:
import shutil
shutil.copy(tts_path, synced_path)
concat_entries.append(synced_path)
playback_cursor += _get_wav_duration(synced_path)
# Write concat list for ffmpeg
with open(concat_list_path, "w") as f:
for entry in concat_entries:
abs_entry = os.path.abspath(entry)
f.write(f"file '{abs_entry}'\n")
# Concatenate all segments (re-encode to normalize sample rates)
cmd = [
"ffmpeg", "-y",
"-f", "concat", "-safe", "0",
"-i", concat_list_path,
"-ar", str(tts_sample_rate),
"-ac", "1",
"-acodec", "pcm_s16le",
output_path,
]
result = subprocess.run(cmd, capture_output=True, text=True)
if result.returncode != 0:
raise RuntimeError(f"ffmpeg concat failed:\n{result.stderr}")
print(f"[s5] Audio sync complete β {output_path} β")
return output_path
|