Loading and composing roles

Every downloadable family has the same canonical loader pair:

• Load(ResolvedModel model, options/threads, cancellationToken) opens already-pulled local files without network access. Obtain the directory/provenance pair from ModelStore.PullAsync, not by guessing cache filenames.
• LoadAsync(ModelStore? store = null, string? model = null, options/threads, CancellationToken cancellationToken = default) pulls, verifies and locally converts missing files before loading. Defaults are the family's declared checkpoint, not a tier-2 default for Speech. Calling a named tier-2 family loader is explicit selection of its documented publisher checkpoint.

Whisper, Moonshine, Qwen3-ASR and Nemotron ASR use typed options; other loaders accept threads, default 1. All family options call this knob Threads, default 1; families permitting 0 document it as runtime-auto. String-directory overloads and graph-path constructors remain for manually converted/custom graphs and require the documented family graph format. They are not needed for catalog loads. Embedded Silero uses its constructor and never downloads.

Rebuild the file pipeline

using SpeechRecognizer recognizer = await WhisperRecognizer.LoadAsync(
    model: "openai/whisper-tiny", cancellationToken: cancellationToken);
using VoiceActivityDetector recognitionVad = new SileroVad(cancellationToken: cancellationToken);
using SpeakerEmbedder embedder = await CampPlusEmbedder.LoadAsync(cancellationToken: cancellationToken);
using VoiceActivityDetector speakerVad = new SileroVad(SileroVadOptions.Diarization, cancellationToken);
using Diarizer diarizer = new ClusteringDiarizer(embedder, speakerVad);
using Aligner qwen = await Qwen3ForcedAligner.LoadAsync(cancellationToken: cancellationToken);
using Aligner ctc = await OmniAsrAligner.LoadAsync(cancellationToken: cancellationToken);
var pipeline = new VadTranscriptionPipeline(recognizer, recognitionVad)
{
    Diarizer = diarizer,
    TimingAligners = [qwen, ctc]
};
Transcript result = pipeline.Transcribe("meeting.wav", "en", cancellationToken);

Namespaces: OneSpeech, .Whisper, .Vad, .Embedders, .Diarization, .Qwen3, .Alignment, .Pipeline. CampPlusEmbedder.LoadAsync(model: "iic/speech_campplus_sv_zh-cn_16k-common") derives the Chinese dimension itself. The example uses explicit English timing/embedding; automatic per-recording language selection and deferred model acquisition are provided by Speech.Transcribe. It eagerly loads both timing candidates, unlike the lazy happy path. It uses default clustering and resegmentation with the diarization-specific VAD profile; no overlap detection is implied.

Roles are disposable owners of their own graph resources. Pipelines, clustering diarizers, live sessions and recognizer decorators borrow their collaborators; disposing a wrapper/session never disposes a borrowed recognizer, detector, aligner or embedder. Dispose sessions before their models and models after all processing has stopped. Dispose is idempotent. Custom subclasses override Dispose(bool) and call the base. Progress uses IProgress<TimeSpan> with processed audio time, never wall time.

SpeechRecognizer.CreateLiveSession(VoiceActivityDetector detector, language) accepts any detector. Silero sessions own independent recurrent state; custom detectors receive one zero-padded 32 ms mono frame at a time. They should be stateless or dedicated to that session. A custom VAD returns intervals, not calibrated probabilities; any positive interval counts as active for endpointing.