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(voice): stage Azure speech output (#1580)

Rudy Mizrahi Celeklib3abbeb40cd1
diff --git a/app/services/voice.py b/app/services/voice.py
index dd654072..a9612ee4 100644
--- a/app/services/voice.py
+++ b/app/services/voice.py
@@ -1790,6 +1790,9 @@ def azure_tts_v2(
return 0
for i in range(3):
+ temporary_audio = None
+ audio_config = None
+ speech_synthesizer = None
try:
logger.info(
f"start, voice name: {voice_name}, rate: {voice_rate}, try: {i + 1}"
@@ -1820,9 +1823,15 @@ def azure_tts_v2(
logger.error("Azure speech key or region is not set")
return None
- audio_config = speechsdk.audio.AudioOutputConfig(
- filename=voice_file, use_default_speaker=True
+ # File and default-speaker output are mutually exclusive SDK modes.
+ # Let each attempt own a private file, never the last good export.
+ ensure_file_path_exists(voice_file)
+ descriptor, temporary_audio = tempfile.mkstemp(
+ prefix=".azure-tts-", suffix=".mp3",
+ dir=os.path.dirname(os.path.abspath(voice_file)),
)
+ os.close(descriptor)
+ audio_config = speechsdk.audio.AudioOutputConfig(filename=temporary_audio)
speech_config = speechsdk.SpeechConfig(
subscription=speech_key, region=service_region
)
@@ -1848,6 +1857,15 @@ def azure_tts_v2(
# 正式生成都会按 WebUI/API 传入的 voice_rate 调整语速。
result = speech_synthesizer.speak_ssml_async(ssml).get()
if result.reason == speechsdk.ResultReason.SynthesizingAudioCompleted:
+ if os.path.getsize(temporary_audio) <= 0:
+ logger.error("Azure completed without audio; preserve the previous export")
+ return None
+ # Release SDK file-output owners before publication, including
+ # on Windows where an open handle can prevent replacement.
+ speech_synthesizer = None
+ audio_config = None
+ os.replace(temporary_audio, voice_file)
+ temporary_audio = None
logger.success(f"azure v2 speech synthesis succeeded: {voice_file}")
return sub_maker
elif result.reason == speechsdk.ResultReason.Canceled:
@@ -1862,6 +1880,18 @@ def azure_tts_v2(
logger.info(f"completed, output file: {voice_file}")
except Exception as e:
logger.error(f"failed, error: {str(e)}")
+ finally:
+ # Exceptions/cancellation can also leave the SDK output owner alive.
+ # Drop all file owners before attempting to remove its staged file.
+ speech_synthesizer = None
+ audio_config = None
+ if temporary_audio is not None:
+ try:
+ os.remove(temporary_audio)
+ except FileNotFoundError:
+ pass
+ except OSError as exc:
+ logger.warning(f"failed to remove staged Azure audio: {exc}")
return None
diff --git a/test/services/test_azure_boundary_duration.py b/test/services/test_azure_boundary_duration.py
index 50ca6e5d..cd864355 100644
--- a/test/services/test_azure_boundary_duration.py
+++ b/test/services/test_azure_boundary_duration.py
@@ -1,4 +1,6 @@
from datetime import timedelta
+from pathlib import Path
+from tempfile import TemporaryDirectory
from types import SimpleNamespace
import unittest
from unittest.mock import Mock, patch
@@ -13,11 +15,20 @@ class TestAzureBoundaryDuration(unittest.TestCase):
captured = {}
synthesizer = Mock()
synthesizer.synthesis_word_boundary.connect.side_effect = lambda callback: captured.update(callback=callback)
+ def configure_output(*, filename):
+ captured["filename"] = filename
+ return Mock()
+
def speak(ssml):
+ # This duration-only SDK fixture must also fulfill the
+ # completed file-output contract before publication.
+ Path(captured["filename"]).write_bytes(b"fixture-audio")
captured["callback"](SimpleNamespace(duration=duration, audio_offset=123, text="word"))
return SimpleNamespace(get=lambda: SimpleNamespace(reason=sdk.ResultReason.SynthesizingAudioCompleted))
synthesizer.speak_ssml_async.side_effect = speak
- with patch.dict(voice.config.azure, {"speech_key": "test-key", "speech_region": "test-region"}), patch.object(sdk.audio, "AudioOutputConfig"), patch.object(sdk, "SpeechConfig"), patch.object(sdk, "SpeechSynthesizer", return_value=synthesizer):
- result = voice.azure_tts_v2("Word.", "en-US-AriaNeural-V2-Female", "unused.mp3")
+ with TemporaryDirectory() as directory, patch.dict(voice.config.azure, {"speech_key": "test-key", "speech_region": "test-region"}), patch.object(sdk.audio, "AudioOutputConfig", side_effect=configure_output), patch.object(sdk, "SpeechConfig"), patch.object(sdk, "SpeechSynthesizer", return_value=synthesizer):
+ output = Path(directory) / "narration.mp3"
+ result = voice.azure_tts_v2("Word.", "en-US-AriaNeural-V2-Female", str(output))
+ self.assertEqual(output.read_bytes(), b"fixture-audio")
self.assertIsNotNone(result)
self.assertEqual(result.offset, [(123, 123 + ticks)])
diff --git a/test/services/test_azure_file_publication.py b/test/services/test_azure_file_publication.py
new file mode 100644
index 00000000..c9f4f5f0
--- /dev/null
+++ b/test/services/test_azure_file_publication.py
@@ -0,0 +1,108 @@
+import shutil
+import subprocess
+import sys
+import weakref
+from datetime import timedelta
+from pathlib import Path
+from types import ModuleType, SimpleNamespace
+from unittest.mock import patch
+
+import pytest
+
+from app.config import config
+from app.services import voice
+
+
[email protected]("outcome,strict,nested", [("completed", True, False), ("completed", True, True), ("completed", False, False), ("cancelled", False, False), ("error", False, False), ("empty", False, False)])
+def test_public_azure_file_output_preserves_last_success(tmp_path, outcome, strict, nested):
+ """Controlled documented SDK file/callback boundary; no real Azure call."""
+ if not shutil.which("ffmpeg"):
+ pytest.skip("native FFmpeg required")
+ complete = subprocess.run(["ffmpeg", "-nostdin", "-v", "error", "-f", "lavfi", "-i", "sine=frequency=880:duration=0.2", "-f", "mp3", "-"], capture_output=True, check=True).stdout
+ output = tmp_path / "nested" / "narration.mp3" if nested else tmp_path / "narration.mp3"
+ previous = b"previous-successful-export"
+ if not nested:
+ output.write_bytes(previous)
+ configured = []
+ open_outputs = set()
+ real_remove = voice.os.remove
+
+ def remove_after_sdk_release(path):
+ assert str(path) not in open_outputs, "SDK still owns the staged output"
+ return real_remove(path)
+ requests = []
+
+ class AudioOutput:
+ def __init__(self, filename=None, use_default_speaker=False):
+ if strict and filename and use_default_speaker:
+ raise ValueError("only one output argument may be selected")
+ if not Path(filename).parent.is_dir():
+ raise OSError("audio output parent does not exist")
+ self.filename = filename
+ self.output = open(filename, "wb")
+ open_outputs.add(str(filename))
+ configured.append(filename)
+
+ def __del__(self):
+ if hasattr(self, "output"):
+ self.output.close()
+ open_outputs.discard(str(self.filename))
+
+ class SpeechConfig:
+ def __init__(self, **_kwargs):
+ pass
+
+ def set_property(self, *_args, **_kwargs):
+ pass
+
+ def set_speech_synthesis_output_format(self, *_args):
+ pass
+
+ class Synthesizer:
+ def __init__(self, audio_config, speech_config):
+ self.audio_config = audio_config
+ self.destination = Path(audio_config.filename)
+ self.callback = None
+ owner = weakref.ref(self)
+ self.synthesis_word_boundary = SimpleNamespace(connect=lambda callback: setattr(owner(), "callback", callback))
+
+ def speak_ssml_async(self, text):
+ requests.append(text)
+
+ def get():
+ self.audio_config.output.write(complete if outcome == "completed" else b"" if outcome == "empty" else complete[:50])
+ self.audio_config.output.flush()
+ if outcome == "error":
+ raise OSError("controlled SDK failure after writing partial bytes")
+ if self.callback and outcome == "completed":
+ self.callback(SimpleNamespace(text="Hello", duration=timedelta(seconds=0.2), audio_offset=0))
+ return SimpleNamespace(reason="cancelled" if outcome == "cancelled" else "completed", cancellation_details=SimpleNamespace(reason="error", error_details="fixture"))
+ return SimpleNamespace(get=get)
+
+ sdk = ModuleType("azure.cognitiveservices.speech")
+ sdk.audio = SimpleNamespace(AudioOutputConfig=AudioOutput)
+ sdk.SpeechConfig = SpeechConfig
+ sdk.SpeechSynthesizer = Synthesizer
+ sdk.SessionEventArgs = object
+ sdk.PropertyId = SimpleNamespace(SpeechServiceResponse_RequestWordBoundary="word")
+ sdk.SpeechSynthesisOutputFormat = SimpleNamespace(Audio48Khz192KBitRateMonoMp3="mp3")
+ sdk.ResultReason = SimpleNamespace(SynthesizingAudioCompleted="completed", Canceled="cancelled")
+ sdk.CancellationReason = SimpleNamespace(Error="error")
+ azure = ModuleType("azure")
+ cognitive = ModuleType("azure.cognitiveservices")
+ azure.cognitiveservices = cognitive
+ cognitive.speech = sdk
+ with patch.object(voice.os, "remove", side_effect=remove_after_sdk_release), patch.dict(sys.modules, {"azure": azure, "azure.cognitiveservices": cognitive, "azure.cognitiveservices.speech": sdk}), patch.dict(config.azure, {"speech_key": "fixture", "speech_region": "fixture"}):
+ result = voice.azure_tts_v2("Hello", "en-US-AriaNeural-V2", str(output))
+ if outcome == "completed":
+ assert result is not None
+ assert result.subs == ["Hello"]
+ assert output.read_bytes() == complete
+ subprocess.run(["ffmpeg", "-nostdin", "-v", "error", "-xerror", "-i", str(output), "-f", "null", "-"], capture_output=True, check=True)
+ else:
+ assert result is None
+ assert output.read_bytes() == previous
+ assert not open_outputs
+ assert all(Path(path) != output for path in configured)
+ assert list(output.parent.iterdir()) == [output]
+ assert len(requests) == (3 if outcome in {"cancelled", "error"} else 1)