👁 Image
FieldCure.DocumentParsers.Audio 0.4.0

FieldCure.DocumentParsers.Audio

Audio transcription support for FieldCure.DocumentParsers.

This package adds an IDocumentParser for .mp3, .wav, .m4a, .ogg, .flac, and .webm files. Audio is converted to 16 kHz mono PCM WAV with NAudio, then transcribed with Whisper.net into timestamped Markdown for RAG pipelines.

Install

dotnet add package FieldCure.DocumentParsers.Audio

Usage

using FieldCure.DocumentParsers;
using FieldCure.DocumentParsers.Audio;

await using var transcriber = DocumentParserFactoryAudioExtensions.AddAudioSupport();

var parser = DocumentParserFactory.GetParser(".mp3")!;
var markdown = parser.ExtractText(File.ReadAllBytes("meeting.mp3"));

For explicit options, instantiate AudioDocumentParser directly:

var parser = new AudioDocumentParser();
var markdown = parser.ExtractText(
 File.ReadAllBytes("meeting.m4a"),
 new AudioExtractionOptions
 {
 SourceExtension = ".m4a",
 Language = "ko",
 ModelSize = WhisperModelSize.Base,
 IncludeTimestamps = true
 });

Set ModelPath to use a pre-downloaded ggml model. Otherwise the default model provider downloads the selected model from Hugging Face on first use and caches it under {UserProfile}/.fieldcure/whisper-models/.

WhisperModelSize.Large resolves to ggml-large-v2 (≈ 3 GB) — large-v3 hallucinates repetition loops on long-form audio, so v2 is the safer default. See RELEASENOTES.DocumentParsers.Audio.md for the v0.2.1 entry behind this choice.

Environment-aware model selection

Instead of hard-coding a ModelSize, let the library pick one based on the detected GPU / RAM / cores of the host:

using FieldCure.DocumentParsers.Audio;

var recommended = WhisperEnvironment.RecommendModelSize(); // QualityBias.Accuracy default
var options = AudioExtractionOptions.Default.WithModelSize(recommended);

// Diagnostic snapshot (CUDA/Vulkan flags, RAM, cores)
var probe = WhisperEnvironment.Probe();
Console.Error.WriteLine(
 $"[Audio] CUDA={probe.CudaAvailable} Vulkan={probe.VulkanAvailable} " +
 $"RAM={probe.SystemRamBytes / (1024L * 1024 * 1024)}GB → {recommended}");

The balanced matrix (used directly by QualityBias.Balanced):

QualityBias.Accuracy (default) shifts the recommendation one tier up — fits batch indexing where transcription latency is acceptable. QualityBias.Speed shifts one tier down for interactive flows.

AudioExtractionOptions.WithModelSize(WhisperModelSize) is a class-friendly substitute for the with expression syntax (the type is a class, not a record), useful when downstream layers want to override only the model size while preserving the rest of the options.

Lifecycle

WhisperTranscriber caches the underlying WhisperFactory for the model path in use, so it holds native resources for as long as it lives. Dispose it once at application shutdown:

// Startup
var transcriber = DocumentParserFactoryAudioExtensions.AddAudioSupport();

// Shutdown
await transcriber.DisposeAsync();

When you construct AudioDocumentParser() with no arguments it owns its transcriber, and await using on the parser disposes it for you.

Runtime selection (CUDA / Vulkan / CPU)

Whisper.net resolves the native runtime through process-global state (Whisper.net.LibraryLoader.RuntimeOptions.RuntimeLibraryOrder). To override the default CUDA → Vulkan → CPU order, set it once at startup rather than per call:

using Whisper.net.LibraryLoader;

// App startup
RuntimeOptions.RuntimeLibraryOrder = [RuntimeLibrary.Cpu];

AudioExtractionOptions.RuntimeLibraryOrder writes to the same global, which is convenient for one-off configuration but is last-writer-wins under concurrent calls.

Notes

• The package is Windows-only because audio decoding uses NAudio's Media Foundation path for several containers. The assembly carries [SupportedOSPlatform("windows")].
• The ggml model is downloaded at runtime; it is not included in the NuGet package.
• For tests or cloud transcription, inject a custom IAudioTranscriber.