Spaces:
Sleeping
Sleeping
File size: 10,639 Bytes
bad74fd | 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 | """
Qwen 3.5 Omni Engine — End-to-end speech-to-speech translation.
Takes English audio in, returns translated audio + transcript out.
No separate ASR/MT/TTS needed — Qwen handles everything in one call.
"""
import os
import base64
import struct
import subprocess
import tempfile
import time
import shutil
import logging
logger = logging.getLogger(__name__)
QWEN_MODEL = "qwen3.5-omni-plus"
QWEN_BASE_URL = "https://dashscope-intl.aliyuncs.com/compatible-mode/v1"
def _get_client():
"""Create OpenAI-compatible client for Qwen Dashscope API."""
from openai import OpenAI
api_key = os.environ.get("DASHSCOPE_API_KEY", "")
if not api_key:
raise RuntimeError(
"DASHSCOPE_API_KEY not set. Add it as a Space secret."
)
return OpenAI(api_key=api_key, base_url=QWEN_BASE_URL)
def _wav_to_base64(wav_path):
"""Read WAV file and return base64 string."""
with open(wav_path, "rb") as f:
return base64.b64encode(f.read()).decode("utf-8")
def _base64_to_wav(b64_data, output_path):
"""Convert raw PCM base64 audio to WAV file (24kHz, mono, 16-bit)."""
audio_bytes = base64.b64decode(b64_data)
sample_rate = 24000
num_channels = 1
bits_per_sample = 16
byte_rate = sample_rate * num_channels * bits_per_sample // 8
block_align = num_channels * bits_per_sample // 8
data_size = len(audio_bytes)
with open(output_path, "wb") as f:
f.write(b"RIFF")
f.write(struct.pack("<I", 36 + data_size))
f.write(b"WAVE")
f.write(b"fmt ")
f.write(struct.pack("<I", 16))
f.write(struct.pack("<H", 1))
f.write(struct.pack("<H", num_channels))
f.write(struct.pack("<I", sample_rate))
f.write(struct.pack("<I", byte_rate))
f.write(struct.pack("<H", block_align))
f.write(struct.pack("<H", bits_per_sample))
f.write(b"data")
f.write(struct.pack("<I", data_size))
f.write(audio_bytes)
def _extract_audio_chunk(video_path, output_wav, start_sec, duration_sec):
"""Extract a chunk of audio from video as 16kHz mono WAV."""
subprocess.run(
["ffmpeg", "-y", "-ss", str(start_sec), "-t", str(duration_sec),
"-i", video_path, "-vn", "-acodec", "pcm_s16le",
"-ar", "16000", "-ac", "1", output_wav],
capture_output=True, check=True,
)
def _get_duration(filepath):
"""Get media file duration in seconds."""
result = subprocess.run(
["ffprobe", "-v", "quiet", "-show_entries", "format=duration",
"-of", "default=noprint_wrappers=1:nokey=1", filepath],
capture_output=True, text=True,
)
return float(result.stdout.strip())
def _concatenate_wavs(wav_files, output_path):
"""Concatenate WAV files using ffmpeg."""
if len(wav_files) == 1:
shutil.copy2(wav_files[0], output_path)
return
list_file = output_path + ".txt"
with open(list_file, "w") as f:
for wav in wav_files:
f.write(f"file '{wav}'\n")
subprocess.run(
["ffmpeg", "-y", "-f", "concat", "-safe", "0",
"-i", list_file, "-c", "copy", output_path],
capture_output=True, check=True,
)
os.remove(list_file)
def _build_system_prompt(language_name):
"""Build Qwen system prompt for a target language."""
return (
f"You are a professional video dubbing translator. You will receive audio in English.\n"
f"Your task:\n"
f"1. Listen carefully to the English speech.\n"
f"2. Translate it into natural, fluent {language_name}.\n"
f"3. Respond ONLY with the {language_name} translation spoken aloud — no English, no commentary,\n"
f" no meta-text, no transliteration. Speak entirely in {language_name}.\n"
f"4. Match the tone, emotion, and pacing of the original speaker as closely as possible.\n"
f"5. If there are pauses or silence in the original audio, maintain similar pacing.\n"
f"6. Translate idioms and cultural references into their {language_name} equivalents.\n"
f"7. Use clear, professional pronunciation suitable for a broad audience."
)
def translate_chunk_qwen(wav_path, voice, language_name, chunk_index=0):
"""
Translate a single audio chunk using Qwen Omni.
Args:
wav_path: Path to input WAV file (English audio)
voice: Qwen voice name (e.g. "Ethan", "Cherry")
language_name: Full language name for the system prompt
chunk_index: For logging
Returns:
(output_wav_path, transcript) or (None, transcript) if no audio
"""
client = _get_client()
audio_b64 = _wav_to_base64(wav_path)
output_wav = wav_path.replace(".wav", f"_qwen_{chunk_index}.wav")
system_prompt = _build_system_prompt(language_name)
user_prompt = f"Translate this English speech into {language_name}. Respond only with the spoken {language_name} translation."
t0 = time.time()
completion = client.chat.completions.create(
model=QWEN_MODEL,
messages=[
{"role": "system", "content": system_prompt},
{
"role": "user",
"content": [
{
"type": "input_audio",
"input_audio": {
"data": f"data:audio/wav;base64,{audio_b64}",
"format": "wav",
},
},
{"type": "text", "text": user_prompt},
],
},
],
modalities=["text", "audio"],
audio={"voice": voice, "format": "wav"},
stream=True,
stream_options={"include_usage": True},
)
audio_chunks = []
transcript_parts = []
for event in completion:
if not event.choices:
continue
delta = event.choices[0].delta
if hasattr(delta, "content") and delta.content:
transcript_parts.append(delta.content)
if hasattr(delta, "audio") and delta.audio:
if isinstance(delta.audio, dict):
if "data" in delta.audio:
audio_chunks.append(delta.audio["data"])
elif hasattr(delta.audio, "data") and delta.audio.data:
audio_chunks.append(delta.audio.data)
transcript = "".join(transcript_parts)
elapsed = time.time() - t0
logger.info(f"Qwen chunk {chunk_index}: {elapsed:.1f}s, transcript={transcript[:60]}")
if audio_chunks:
full_audio_b64 = "".join(audio_chunks)
_base64_to_wav(full_audio_b64, output_wav)
return output_wav, transcript
return None, transcript
def dub_video_qwen(video_path, language_name, voice="Ethan", chunk_seconds=120, progress_fn=None):
"""
Full video dubbing pipeline using Qwen Omni.
Splits video into chunks, translates each chunk via Qwen API,
concatenates results, and muxes back onto video.
Args:
video_path: Path to input video
language_name: Full language name (e.g. "French", "Arabic")
voice: Qwen voice name
chunk_seconds: Audio chunk duration for API calls
progress_fn: Optional gradio progress callback
Returns:
(output_video_path, log_text)
"""
tmp_dir = tempfile.mkdtemp(prefix=f"qwen_dub_")
log = []
try:
# Duration
if progress_fn:
progress_fn(0.05, desc="Analyzing video...")
total_duration = _get_duration(video_path)
log.append(f"**Video:** {total_duration:.1f}s")
log.append(f"**Engine:** Qwen 3.5 Omni")
log.append(f"**Voice:** {voice}")
log.append(f"**Language:** {language_name}")
if total_duration > 3600:
return None, "Video longer than 1 hour — please use a shorter clip."
# Split into chunks
if progress_fn:
progress_fn(0.1, desc="Extracting audio chunks...")
num_chunks = max(1, int(total_duration // chunk_seconds) + (1 if total_duration % chunk_seconds > 0 else 0))
log.append(f"**Chunks:** {num_chunks} ({chunk_seconds}s each)")
input_chunks = []
for i in range(num_chunks):
start = i * chunk_seconds
duration = min(chunk_seconds, total_duration - start)
chunk_path = os.path.join(tmp_dir, f"chunk_{i:03d}.wav")
_extract_audio_chunk(video_path, chunk_path, start, duration)
input_chunks.append(chunk_path)
# Translate each chunk
output_chunks = []
all_transcripts = []
for i, chunk_path in enumerate(input_chunks):
if progress_fn:
frac = 0.15 + 0.7 * (i / num_chunks)
progress_fn(frac, desc=f"Translating chunk {i+1}/{num_chunks}...")
result_path, transcript = translate_chunk_qwen(
chunk_path, voice, language_name, i
)
if transcript:
all_transcripts.append(f"**[{i+1}]** {transcript}")
if result_path:
output_chunks.append(result_path)
else:
# Silence fallback
duration = _get_duration(chunk_path)
silence_path = os.path.join(tmp_dir, f"silence_{i:03d}.wav")
subprocess.run(
["ffmpeg", "-y", "-f", "lavfi",
"-i", "anullsrc=r=24000:cl=mono",
"-t", str(duration), "-acodec", "pcm_s16le", silence_path],
capture_output=True, check=True,
)
output_chunks.append(silence_path)
# Concatenate
if progress_fn:
progress_fn(0.88, desc="Assembling audio...")
full_audio = os.path.join(tmp_dir, "full_dubbed.wav")
_concatenate_wavs(output_chunks, full_audio)
# Mux onto video
if progress_fn:
progress_fn(0.93, desc="Combining audio and video...")
output_video = os.path.join(tmp_dir, "dubbed_output.mp4")
subprocess.run(
["ffmpeg", "-y", "-i", video_path, "-i", full_audio,
"-c:v", "copy", "-map", "0:v:0", "-map", "1:a:0",
"-shortest", output_video],
capture_output=True, check=True,
)
if progress_fn:
progress_fn(1.0, desc="Done!")
log.append(f"\n**Transcript:**")
log.extend(all_transcripts)
return output_video, "\n".join(log)
except Exception as e:
logger.exception("Qwen dubbing failed")
shutil.rmtree(tmp_dir, ignore_errors=True)
return None, f"Error: {str(e)}"
|