Skip to content

利用 AI 大模型和自动化工作流,根据主题或关键词一键生成高清短视频。Generate HD short videos from a topic or keyword with an automated AI workflow.

fix(tts): time out FFmpeg narration (#1472)

Rudy Mizrahi Celeklicce2da724e9f
diff --git a/app/services/voice.py b/app/services/voice.py
index b5253755..53ce9148 100644
--- a/app/services/voice.py
+++ b/app/services/voice.py
@@ -85,6 +85,7 @@ VOXCPM_REFERENCE_AUDIO_MAX_WAV_BYTES = 5 * 1024 * 1024
VOXCPM_REFERENCE_AUDIO_CONVERSION_TIMEOUT_SECONDS = 15
VOXCPM_REFERENCE_AUDIO_MAX_DURATION_SECONDS = 120
VOXCPM_REFERENCE_AUDIO_FILE_TYPES = ("wav", "mp3", "m4a", "aac", "ogg", "flac")
+_DEFAULT_TTS_FFMPEG_TIMEOUT_SECONDS = 600
_VOXCPM_NON_RETRYABLE_STATUS_CODES = {400, 401, 403, 404, 422}
_VOXCPM_RETRY_DELAY_SECONDS = (1.0, 2.0)
NO_VOICE_NAME = "no-voice"
@@ -557,6 +558,9 @@ def generate_silent_audio(duration_seconds: float, output_file: str) -> bool:
ffmpeg_binary = utils.get_ffmpeg_binary()
command = [
ffmpeg_binary,
+ "-nostdin",
+ "-v",
+ "error",
"-y",
"-f",
"lavfi",
@@ -568,33 +572,12 @@ def generate_silent_audio(duration_seconds: float, output_file: str) -> bool:
"libmp3lame",
"-q:a",
"4",
- output_file,
]
logger.info(
f"generating silent audio for no-voice mode, duration: {duration_seconds:.2f}s"
)
- result = subprocess.run(
- command,
- capture_output=True,
- text=True,
- encoding="utf-8",
- errors="replace",
- check=False,
- )
- if result.returncode != 0:
- logger.error(
- "failed to generate silent audio: "
- f"{(result.stderr or result.stdout or '').strip()}"
- )
- return False
- if not os.path.exists(output_file) or os.path.getsize(output_file) <= 0:
- logger.error(
- "silent audio output file is missing or empty, "
- f"file: {output_file}, duration: {duration_seconds:.2f}s"
- )
- return False
- return True
+ return _publish_tts_ffmpeg_output(command, output_file, "silent narration encode")
def _single_tts(
@@ -731,6 +714,69 @@ def _single_tts(
return azure_tts_v1(text, voice_name, voice_rate, voice_file)
+def _run_tts_ffmpeg(command: list[str], stage: str):
+ """Bound TTS FFmpeg work and reap the child on timeout."""
+ raw_timeout = config.app.get(
+ "ffmpeg_tts_timeout_seconds",
+ _DEFAULT_TTS_FFMPEG_TIMEOUT_SECONDS,
+ )
+ try:
+ timeout = float(raw_timeout)
+ except (TypeError, ValueError) as exc:
+ raise ValueError("ffmpeg_tts_timeout_seconds must be positive") from exc
+ if not math.isfinite(timeout) or timeout <= 0:
+ raise ValueError("ffmpeg_tts_timeout_seconds must be positive")
+
+ try:
+ # subprocess.run kills and waits for the child on timeout. Do not allow
+ # FFmpeg to wait for terminal input in an unattended generation task.
+ return subprocess.run(
+ command,
+ capture_output=True,
+ text=True,
+ encoding="utf-8",
+ errors="replace",
+ check=False,
+ timeout=timeout,
+ )
+ except subprocess.TimeoutExpired:
+ logger.error(f"FFmpeg {stage} timed out after {timeout:g} seconds")
+ except OSError as exc:
+ logger.error(f"failed to start FFmpeg {stage}: {exc}")
+ return None
+
+
+def _publish_tts_ffmpeg_output(
+ command: list[str], output_file: str, stage: str
+) -> bool:
+ """Encode beside the destination and replace only after a complete output."""
+ output_dir = os.path.dirname(output_file) or "."
+ with tempfile.TemporaryDirectory(
+ prefix=".mpt-tts-audio-", dir=output_dir
+ ) as publish_temp:
+ staged_output = os.path.join(
+ publish_temp, f"narration{os.path.splitext(output_file)[1] or '.mp3'}"
+ )
+ result = _run_tts_ffmpeg([*command, staged_output], stage)
+ if result is None:
+ return False
+ if result.returncode != 0:
+ logger.error(
+ f"FFmpeg {stage} failed: "
+ f"{(result.stderr or result.stdout or '').strip()}"
+ )
+ return False
+ if not os.path.exists(staged_output) or os.path.getsize(staged_output) == 0:
+ logger.error(f"FFmpeg {stage} produced no narration audio")
+ return False
+ try:
+ os.replace(staged_output, output_file)
+ except OSError as exc:
+ logger.error(f"failed to publish {stage} audio: {exc}")
+ return False
+ return True
+
+
def _concat_audio_files(audio_files: list[str], output_file: str) -> bool:
"""
使用 PCM 解码与统一重编码合并多个音频分段。
@@ -791,6 +837,9 @@ def _concat_audio_files(audio_files: list[str], output_file: str) -> bool:
pcm_wav = os.path.join(concat_temp, f"chunk_{idx}.wav")
cmd = [
ffmpeg_binary,
+ "-nostdin",
+ "-v",
+ "error",
"-y",
"-i",
f,
@@ -803,7 +852,9 @@ def _concat_audio_files(audio_files: list[str], output_file: str) -> bool:
"pcm_s16le",
pcm_wav,
]
- res = subprocess.run(cmd, capture_output=True, text=True, encoding="utf-8", errors="replace", check=False)
+ res = _run_tts_ffmpeg(cmd, "pause audio decode")
+ if res is None:
+ return False
if res.returncode == 0 and os.path.exists(pcm_wav):
try:
with wave.open(pcm_wav, "rb") as wf:
@@ -824,6 +875,9 @@ def _concat_audio_files(audio_files: list[str], output_file: str) -> bool:
command = [
ffmpeg_binary,
+ "-nostdin",
+ "-v",
+ "error",
"-y",
"-i",
temp_combined_wav,
@@ -831,16 +885,10 @@ def _concat_audio_files(audio_files: list[str], output_file: str) -> bool:
"libmp3lame",
"-q:a",
"4",
- output_file,
]
- result = subprocess.run(command, capture_output=True, text=True, encoding="utf-8", errors="replace", check=False)
- if result.returncode != 0:
- logger.error(
- "failed to encode concatenated audio to mp3: "
- f"{(result.stderr or result.stdout or '').strip()}"
- )
- return False
- return os.path.exists(output_file) and os.path.getsize(output_file) > 0
+ return _publish_tts_ffmpeg_output(
+ command, output_file, "pause audio encode"
+ )
def _tts_with_pauses(
@@ -933,6 +981,9 @@ def _tts_with_pauses(
ffmpeg_binary = utils.get_ffmpeg_binary()
cmd = [
ffmpeg_binary,
+ "-nostdin",
+ "-v",
+ "error",
"-y",
"-i",
chunk_audio_file,
@@ -945,7 +996,9 @@ def _tts_with_pauses(
"pcm_s16le",
chunk_wav,
]
- res = subprocess.run(cmd, capture_output=True, text=True, encoding="utf-8", errors="replace", check=False)
+ res = _run_tts_ffmpeg(cmd, "speech chunk decode")
+ if res is None:
+ return None
if res.returncode != 0 or not os.path.exists(chunk_wav) or os.path.getsize(chunk_wav) == 0:
logger.error(
f"failed to decode speech chunk audio to PCM WAV: {speech_text[:50]}, "
diff --git a/config.example.toml b/config.example.toml
index d7efbfb5..ad649350 100644
--- a/config.example.toml
+++ b/config.example.toml
@@ -445,6 +445,12 @@ subtitle_provider = "edge"
# is stopped and the task reports a failure instead of staying at 50% forever.
# ffmpeg_concat_timeout_seconds = 3600
+# No-voice MP3 and pause-tag narration use FFmpeg. Bound each invocation so a
+# stuck conversion fails the task rather than waiting indefinitely (default:
+# 600 seconds per invocation). Increase this on unusually slow machines or for
+# very long narration chunks.
+# ffmpeg_tts_timeout_seconds = 600
+
# Leave unset to follow the app default (libx264). Hardware encoders are optional;
# unsupported encoders automatically fall back to libx264.
# Available values: "libx264", "h264_nvenc", "h264_amf", "h264_qsv",
diff --git a/test/services/test_voice.py b/test/services/test_voice.py
index 6655b83d..5ec52be8 100644
--- a/test/services/test_voice.py
+++ b/test/services/test_voice.py
@@ -2,6 +2,7 @@ import asyncio
import base64
import os
import shutil
+import subprocess
import unittest
import sys
import tempfile
@@ -187,6 +188,20 @@ class TestVoiceService(unittest.TestCase):
self.assertFalse(vs.generate_silent_audio(3.0, voice_file))
+ def test_no_voice_ffmpeg_timeout_preserves_previous_output(self):
+ def timed_out(command, **kwargs):
+ self.assertIn("-nostdin", command)
+ self.assertGreater(kwargs.get("timeout", 0), 0)
+ Path(command[-1]).write_bytes(b"partial-silence")
+ raise subprocess.TimeoutExpired(command, kwargs["timeout"])
+
+ with tempfile.TemporaryDirectory() as temp_dir:
+ output = Path(temp_dir) / "silent.mp3"
+ output.write_bytes(b"previous-success")
+ with patch.object(vs.subprocess, "run", side_effect=timed_out):
+ self.assertFalse(vs.generate_silent_audio(3.0, str(output)))
+ self.assertEqual(output.read_bytes(), b"previous-success")
+
def test_empty_voice_name_does_not_enable_no_voice_mode(self):
"""
空 voice 通常意味着配置缺失或接口参数错误,不能自动切到无配音模式。
@@ -1876,6 +1891,42 @@ class TestElevenLabsVoice(unittest.TestCase):
self.assertTrue(read_sizes)
self.assertLessEqual(max(read_sizes), 8192)
+ def test_pause_audio_decode_timeout_fails_without_publishing(self):
+ def timed_out(command, **kwargs):
+ self.assertIn("-nostdin", command)
+ self.assertGreater(kwargs.get("timeout", 0), 0)
+ raise subprocess.TimeoutExpired(command, kwargs["timeout"])
+
+ with tempfile.TemporaryDirectory() as temp_dir:
+ inputs = [str(Path(temp_dir) / f"chunk-{index}.mp3") for index in range(2)]
+ for input_path in inputs:
+ Path(input_path).write_bytes(b"compressed-audio")
+ output = str(Path(temp_dir) / "voice.mp3")
+ with patch.object(vs.subprocess, "run", side_effect=timed_out):
+ self.assertFalse(vs._concat_audio_files(inputs, output))
+ self.assertFalse(Path(output).exists())
+
+ def test_pause_audio_encode_timeout_preserves_previous_output(self):
+ def timed_out(command, **kwargs):
+ self.assertIn("-nostdin", command)
+ self.assertGreater(kwargs.get("timeout", 0), 0)
+ Path(command[-1]).write_bytes(b"partial-output")
+ raise subprocess.TimeoutExpired(command, kwargs["timeout"])
+
+ with tempfile.TemporaryDirectory() as temp_dir:
+ inputs = [str(Path(temp_dir) / f"chunk-{index}.wav") for index in range(2)]
+ for input_path in inputs:
+ with vs.wave.open(input_path, "wb") as writer:
+ writer.setnchannels(1)
+ writer.setsampwidth(2)
+ writer.setframerate(24000)
+ writer.writeframes(b"\0\0" * 20)
+ output = Path(temp_dir) / "voice.mp3"
+ output.write_bytes(b"previous-success")
+ with patch.object(vs.subprocess, "run", side_effect=timed_out):
+ self.assertFalse(vs._concat_audio_files(inputs, str(output)))
+ self.assertEqual(output.read_bytes(), b"previous-success")
+
def test_tts_with_pauses_shifts_submaker_timeline(self):
"""测试包含停顿标签时,SubMaker 时间轴和音频拼接正确偏移。"""
from edge_tts.srt_composer import Subtitle
@@ -2371,6 +2422,31 @@ class TestElevenLabsVoice(unittest.TestCase):
)
self.assertIsNone(result)
+ def test_tts_with_pauses_decode_timeout_returns_failure(self):
+ def fake_single_tts(text, voice_name, voice_rate, voice_file, voice_volume=1.0):
+ Path(voice_file).write_bytes(b"synthesized-audio")
+ return vs.SubMaker()
+
+ def timed_out(command, **kwargs):
+ self.assertIn("-nostdin", command)
+ self.assertGreater(kwargs.get("timeout", 0), 0)
+ raise subprocess.TimeoutExpired(command, kwargs["timeout"])
+
+ with tempfile.TemporaryDirectory() as temp_dir:
+ output = Path(temp_dir) / "narration.mp3"
+ with (
+ patch.object(vs, "_single_tts", side_effect=fake_single_tts),
+ patch.object(vs.subprocess, "run", side_effect=timed_out),
+ ):
+ result = vs._tts_with_pauses(
+ text="Hello [pause: 1s] world",
+ voice_name="zh-CN-XiaoxiaoNeural",
+ voice_rate=1.0,
+ voice_file=str(output),
+ )
+ self.assertIsNone(result)
+ self.assertFalse(output.exists())
+
def test_tts_passes_original_text_unchanged_without_pauses(self):
"""测试无停顿标签时,tts 将原始文本原样直通给 _single_tts,不执行正则替换或清洗。"""
original_text = " Leading and trailing spaces, [regular bracket] and punctuation! \nNew line here. "