Warm up pyannote and FunASR runtimes
PR and Push Build/Test / build-and-test (push) Successful in 6m48s

This commit is contained in:
2026-05-28 13:05:27 +02:00
parent 7ff93b73b3
commit c46c587e65
14 changed files with 696 additions and 20 deletions
+2 -1
View File
@@ -181,7 +181,7 @@ public sealed class AzureSpeechOptions
public string KeyEnv { get; set; } = "AZURE_SPEECH_KEY";
public TimeSpan RecognitionStopTimeout { get; set; } = TimeSpan.FromSeconds(10);
public TimeSpan RecognitionStopTimeout { get; set; } = TimeSpan.FromMinutes(3);
public bool DiarizeIntermediateResults { get; set; } = true;
@@ -321,6 +321,7 @@ public sealed class SpeakerIdentityPyannoteValidationOptions
public PyannoteDiarizationOptions Diarization { get; set; } = new()
{
Enabled = true,
CommandTimeout = TimeSpan.FromHours(1),
AlignmentMode = PyannoteAlignmentMode.PyannoteTurns
};
}
+1
View File
@@ -89,6 +89,7 @@ builder.Services.AddSingleton<MeetingRecordingCoordinator>();
builder.Services.AddHostedService<TemporaryRecordingCleanupHostedService>();
builder.Services.AddHostedService<SpeakerIdentityDatabaseInitializer>();
builder.Services.AddHostedService<SpeechRecognitionPipelineHostedService>();
builder.Services.AddHostedService<PyannoteDiarizationWarmupHostedService>();
#if WINDOWS
builder.Services.AddHostedService<GlobalHotkeyService>();
builder.Services.AddHostedService<UnoTaskbarIconService>();
@@ -0,0 +1,125 @@
using MeetingAssistant.LaunchProfiles;
namespace MeetingAssistant.Transcription;
public sealed class PyannoteDiarizationWarmupHostedService : IHostedService
{
private readonly PyannoteTranscriptFinalizer finalizer;
private readonly ILaunchProfileOptionsProvider launchProfiles;
private readonly ILogger<PyannoteDiarizationWarmupHostedService> logger;
private CancellationTokenSource? startupCancellation;
private Task? startupTask;
public PyannoteDiarizationWarmupHostedService(
PyannoteTranscriptFinalizer finalizer,
ILaunchProfileOptionsProvider launchProfiles,
ILogger<PyannoteDiarizationWarmupHostedService> logger)
{
this.finalizer = finalizer;
this.launchProfiles = launchProfiles;
this.logger = logger;
}
public Task StartAsync(CancellationToken cancellationToken)
{
var cancellation = CancellationTokenSource.CreateLinkedTokenSource(cancellationToken);
startupCancellation = cancellation;
startupTask = Task.Run(
() => WarmUpEnabledRuntimesAsync(cancellation.Token),
CancellationToken.None);
return Task.CompletedTask;
}
public async Task StopAsync(CancellationToken cancellationToken)
{
var cancellation = startupCancellation;
var task = startupTask;
if (cancellation is null || task is null)
{
return;
}
startupCancellation = null;
startupTask = null;
await cancellation.CancelAsync();
try
{
await task.WaitAsync(cancellationToken);
}
catch (OperationCanceledException)
{
}
finally
{
cancellation.Dispose();
}
}
private async Task WarmUpEnabledRuntimesAsync(CancellationToken cancellationToken)
{
foreach (var diarization in GetEnabledDiarizationOptions())
{
try
{
logger.LogInformation(
"Starting pyannote warm-up for image {Image} and model {Model}",
diarization.Image,
diarization.Model);
await finalizer.WarmUpAsync(diarization, cancellationToken);
logger.LogInformation(
"Finished pyannote warm-up for image {Image} and model {Model}",
diarization.Image,
diarization.Model);
}
catch (OperationCanceledException) when (cancellationToken.IsCancellationRequested)
{
logger.LogInformation("Pyannote warm-up was cancelled during application shutdown");
throw;
}
catch (Exception exception)
{
logger.LogError(
exception,
"Pyannote warm-up failed for image {Image} and model {Model}; diarization can still retry on demand",
diarization.Image,
diarization.Model);
}
}
}
private IEnumerable<PyannoteDiarizationOptions> GetEnabledDiarizationOptions()
{
return launchProfiles.GetProfiles()
.SelectMany(profile => GetEnabledDiarizationOptions(profile.Options))
.DistinctBy(CreateWarmUpKey);
}
private static IEnumerable<PyannoteDiarizationOptions> GetEnabledDiarizationOptions(
MeetingAssistantOptions options)
{
if (options.Recording.TranscriptionProvider.Equals("whisper-local", StringComparison.OrdinalIgnoreCase) &&
options.WhisperLocal.Diarization.Enabled)
{
yield return options.WhisperLocal.Diarization;
}
if (options.SpeakerIdentification.PyannoteValidation.Enabled &&
options.SpeakerIdentification.PyannoteValidation.Diarization.Enabled)
{
yield return options.SpeakerIdentification.PyannoteValidation.Diarization;
}
}
private static string CreateWarmUpKey(PyannoteDiarizationOptions diarization)
{
return string.Join(
'\u001f',
diarization.DockerCommand,
diarization.Image,
diarization.Model,
VaultPath.Resolve(diarization.ModelsFolder),
diarization.Token,
diarization.TokenEnv,
diarization.BuildImage.ToString());
}
}
@@ -82,6 +82,58 @@ public sealed class PyannoteTranscriptFinalizer
return segments;
}
public async Task WarmUpAsync(
PyannoteDiarizationOptions diarization,
CancellationToken cancellationToken)
{
if (!diarization.Enabled)
{
return;
}
var token = ResolveToken(diarization);
if (string.IsNullOrWhiteSpace(token))
{
logger.LogWarning(
"Pyannote warm-up is enabled but no Hugging Face token is configured directly or via {TokenEnv}",
diarization.TokenEnv);
return;
}
var modelsFolder = VaultPath.Resolve(diarization.ModelsFolder);
Directory.CreateDirectory(modelsFolder);
using var timeoutSource = diarization.CommandTimeout > TimeSpan.Zero
? new CancellationTokenSource(diarization.CommandTimeout)
: null;
using var linkedSource = timeoutSource is null
? null
: CancellationTokenSource.CreateLinkedTokenSource(cancellationToken, timeoutSource.Token);
try
{
if (diarization.BuildImage)
{
await EnsureDockerImageAsync(diarization, modelsFolder, linkedSource?.Token ?? cancellationToken);
}
var result = await commandRunner.RunAsync(
diarization.DockerCommand,
BuildWarmUpDockerArguments(diarization, modelsFolder),
linkedSource?.Token ?? cancellationToken,
new Dictionary<string, string> { ["HF_TOKEN"] = token });
if (result.ExitCode != 0)
{
throw new InvalidOperationException(
$"pyannote warm-up failed with exit code {result.ExitCode}: {result.StandardError}");
}
}
catch (OperationCanceledException) when (!cancellationToken.IsCancellationRequested && timeoutSource?.IsCancellationRequested == true)
{
throw new TimeoutException(
$"pyannote warm-up timed out after {diarization.CommandTimeout}.");
}
}
private async Task<CommandResult> RunDiarizationAsync(
string fullAudioPath,
string token,
@@ -189,6 +241,33 @@ public sealed class PyannoteTranscriptFinalizer
];
}
private string[] BuildWarmUpDockerArguments(
PyannoteDiarizationOptions diarization,
string modelsFolder)
{
return
[
"run",
"--rm",
"-e",
"HF_TOKEN",
"-v",
$"{modelsFolder}:/workspace/cache",
"-e",
"HF_HOME=/workspace/cache/huggingface",
"-e",
"XDG_CACHE_HOME=/workspace/cache",
"-e",
"PIP_CACHE_DIR=/workspace/cache/pip",
"-e",
"TORCH_HOME=/workspace/cache/torch",
diarization.Image,
"sh",
"-lc",
BuildWarmUpPythonCommand(diarization)
];
}
private static string BuildPythonCommand(
PyannoteDiarizationOptions diarization,
SpeechRecognitionPipelineOptions pipelineOptions)
@@ -217,6 +296,21 @@ public sealed class PyannoteTranscriptFinalizer
+ "python /tmp/meeting_assistant_pyannote.py";
}
private static string BuildWarmUpPythonCommand(PyannoteDiarizationOptions diarization)
{
var model = diarization.Model;
return
"cat > /tmp/meeting_assistant_pyannote_warmup.py <<'PY'\n"
+ "import os\n"
+ "import torch\n"
+ "from pyannote.audio import Pipeline\n"
+ $"pipeline = Pipeline.from_pretrained({JsonSerializer.Serialize(model)}, token=os.environ.get('HF_TOKEN'))\n"
+ "pipeline.to(torch.device('cpu'))\n"
+ "print('pyannote warm-up complete')\n"
+ "PY\n"
+ "python /tmp/meeting_assistant_pyannote_warmup.py";
}
private static string BuildPipelineInvocation(SpeechRecognitionPipelineOptions pipelineOptions)
{
return pipelineOptions.NumSpeakers is > 0
@@ -1,27 +1,41 @@
using MeetingAssistant.LaunchProfiles;
namespace MeetingAssistant.Transcription;
public sealed class SpeechRecognitionPipelineHostedService : IHostedService
{
private readonly ISpeechRecognitionPipelineFactory pipelineFactory;
private readonly ILaunchProfileOptionsProvider? launchProfiles;
private readonly ILogger<SpeechRecognitionPipelineHostedService> logger;
private CancellationTokenSource? startupCancellation;
private Task? startupTask;
private ISpeechRecognitionPipeline? startupPipeline;
private List<ISpeechRecognitionPipeline> startupPipelines = [];
public SpeechRecognitionPipelineHostedService(
ISpeechRecognitionPipelineFactory pipelineFactory,
ILogger<SpeechRecognitionPipelineHostedService> logger)
: this(pipelineFactory, null, logger)
{
}
public SpeechRecognitionPipelineHostedService(
ISpeechRecognitionPipelineFactory pipelineFactory,
ILaunchProfileOptionsProvider? launchProfiles,
ILogger<SpeechRecognitionPipelineHostedService> logger)
{
this.pipelineFactory = pipelineFactory;
this.launchProfiles = launchProfiles;
this.logger = logger;
}
public Task StartAsync(CancellationToken cancellationToken)
{
startupCancellation = CancellationTokenSource.CreateLinkedTokenSource(cancellationToken);
startupPipeline = pipelineFactory.Create();
var cancellation = CancellationTokenSource.CreateLinkedTokenSource(cancellationToken);
var pipelines = CreateStartupPipelines();
startupCancellation = cancellation;
startupPipelines = pipelines;
startupTask = Task.Run(
() => WarmUpPipelineAsync(startupPipeline, startupCancellation.Token),
() => WarmUpPipelinesAsync(pipelines, cancellation.Token),
CancellationToken.None);
return Task.CompletedTask;
}
@@ -30,15 +44,15 @@ public sealed class SpeechRecognitionPipelineHostedService : IHostedService
{
var cancellation = startupCancellation;
var task = startupTask;
var pipeline = startupPipeline;
if (cancellation is null || task is null || pipeline is null)
var pipelines = startupPipelines;
if (cancellation is null || task is null)
{
return;
}
startupCancellation = null;
startupTask = null;
startupPipeline = null;
startupPipelines = [];
await cancellation.CancelAsync();
try
{
@@ -49,11 +63,43 @@ public sealed class SpeechRecognitionPipelineHostedService : IHostedService
}
finally
{
await pipeline.DisposeAsync();
foreach (var pipeline in pipelines)
{
await pipeline.DisposeAsync();
}
cancellation.Dispose();
}
}
private List<ISpeechRecognitionPipeline> CreateStartupPipelines()
{
if (launchProfiles is null)
{
return [pipelineFactory.Create()];
}
return launchProfiles.GetProfiles()
.Where(ShouldWarmUpProfile)
.Select(profile => pipelineFactory.Create(profile.Name))
.ToList();
}
private static bool ShouldWarmUpProfile(LaunchProfile profile)
{
return profile.Name.Equals(
ConfigurationLaunchProfileOptionsProvider.DefaultProfileName,
StringComparison.OrdinalIgnoreCase) ||
profile.Options.Recording.TranscriptionProvider.Equals("funasr", StringComparison.OrdinalIgnoreCase);
}
private async Task WarmUpPipelinesAsync(
IReadOnlyList<ISpeechRecognitionPipeline> pipelines,
CancellationToken cancellationToken)
{
await Task.WhenAll(pipelines.Select(pipeline => WarmUpPipelineAsync(pipeline, cancellationToken)));
}
private async Task WarmUpPipelineAsync(
ISpeechRecognitionPipeline pipeline,
CancellationToken cancellationToken)
+2 -2
View File
@@ -89,7 +89,7 @@
"de-DE"
],
"KeyEnv": "AZURE_SPEECH_KEY",
"RecognitionStopTimeout": "00:00:10",
"RecognitionStopTimeout": "00:03:00",
"DiarizeIntermediateResults": true,
"PostProcessingOption": "",
"PhraseListWeight": 1.5
@@ -124,7 +124,7 @@
"AnnotationSource": "SpeakerDiarization",
"AlignmentMode": "PyannoteTurns",
"TokenEnv": "HF_TOKEN",
"CommandTimeout": "00:30:00"
"CommandTimeout": "01:00:00"
}
}
},