Speech recognition for .NET

OneSpeech transcribes audio, times words, labels speakers and supports live sessions in managed .NET 10. Open-weight models run on OneModel; OneAudio decodes and resamples audio. No native inference runtime is required.

OneSpeech is unaffiliated with other products or companies named OneSpeech.

See measured results and limitations before choosing a model. Licensed under MIT OR Apache-2.0.

Transcribe, stream and align

using OneSpeech;
 
Transcript transcript = Speech.Transcribe("meeting.m4a", language: "en");
Console.WriteLine(transcript.Text);
File.WriteAllText("meeting.json", transcript.ToJson());
 
await foreach (Segment segment in Speech.TranscribeStream("lecture.wav"))
    Console.WriteLine(segment.Text);
 
Transcript aligned = Speech.Align("short.wav", "The words that were spoken.", language: "en");

Transcribe returns a timed, diarized, language-tagged Transcript: inspect Words, Segments, Turns and Models. Failed word timing is null, never invented. A speaker label is anonymous and local to the document, not a person's identity. TranscribeStream yields final segments in bounded chunks; it does not perform whole-file speaker assignment. Standalone Align accepts at most 60 seconds. All verbs take an optional final cancellationToken; path and encoded-stream I/O have *Async forms. CPU inference remains synchronous within those calls. Language tags are case-insensitive: en-US, EN, eng and English resolve to en; unsupported hints throw CapabilityNotSupportedException.

Speech.Listen consumes caller-supplied mono 16 kHz PCM as an IAsyncEnumerable<AudioBuffer>. It emits replacing partials, immutable finals, endpoints and speaker turns on one audio clock. It does not capture a microphone. See live sessions.

For speaker recognition, use Speech.Enroll("alice", "enrollment.wav"), Speech.Identify("query.wav") and Speech.Verify("first.wav", "second.wav"); encoded streams have *Async twins. Enrollments live only in process memory: Speech.ClearSpeakers() removes them and releases speaker models. File transcription adds matching enrolled names to speaker labels without changing their document-local IDs; unknown clusters keep anonymous labels. Obtain consent and check local biometric law before using these verbs. See speaker recognition for calibrated thresholds, limitations and the low-level registry.

Overlapping speech

The default clustering diarizer does not detect overlapping speech. For an overlap-aware upgrade (up to eight speakers), load the named tier-2 diarizer in one call: await NemotronDiarizer.LoadAsync(), then pass it to your transcription pipeline. This does not separate simultaneous voices. See the complete replacement example and Nemotron guide. It is never downloaded as a default.

First-use downloads and offline preparation

The model argument selects a role, not the entire pipeline. Transcription also needs a speaker embedder and, without native timings, companion aligners. Even Whisper tiny can require several GB on first use. Silero VAD is embedded. See the per-verb download table for exact IDs, sizes and conditional dependencies.

await Speech.PrefetchAsync(language: "en", model: "openai/whisper-tiny"); prepares the complete companion set, including timing fallbacks, without running inference. For live use, pass model: "moonshine-ai/moonshine-streaming-tiny". Every downloadable family exposes Load(ResolvedModel, …) and LoadAsync(store, model, …); composition examples use documented role types.

Contextual biasing

Biasing is a low-level request; Speech.* stays unchanged. A loaded recognizer returns a borrowed wrapper with its own immutable phrase list:

using var model = OneSpeech.Whisper.WhisperRecognizer.Load(modelDirectory);
SpeechRecognizer biased = model.WithPhrases([new Phrase("Acme Photonics"), new Phrase("Marisol")]);
Transcript result = biased.Transcribe(audio, "en", cancellationToken);
// The same biased recognizer can be passed to VadTranscriptionPipeline.

Keep the owning model alive until the wrappers and their sessions are finished. Wrappers do not own or dispose it; the model's concurrency limits still apply. Calling WithPhrases again replaces the list; an empty list returns the original unbiased recognizer. The original recognizer is never modified.

Whisper conditions each window with ordinary prompt tokens. Qwen3-ASR appends phrases to its existing system context. These prompt methods refuse explicit numeric weights rather than pretending to apply a score bonus. Parakeet TDT uses token-level shallow fusion; new Phrase("Acme", 0.5) supplies a natural-log bonus per token (null defaults to 1). Custom CTC/transducer graphs use the same fusion after the caller supplies TokenVocabulary.WithPhraseEncoder for their exact tokenizer. CTC biasing uses prefix beam search, even if the unbiased recognizer is greedy. Moonshine does not support biasing: a nonempty request throws CapabilityNotSupportedException, as declared by Capabilities.Biasing.

Biasing is not a spelling replacement or a guarantee that a term will be spoken. It can introduce false alarms. Whisper prompt conditioning in particular can substantially increase overall errors and repeat absent terms even when recall of requested terms improves; it is not a safe default. Check both phrase recall and overall WER, with absent-phrase controls, on your audio. See contextual scoring and bounds for prompt budgets, weights, decoder semantics and the repeatable evaluation.

Packages

Models not supplied by a package download on first use; CPU runtime graphs are built and verified in the per-user cache. An offline model package covers only that model, not every role used by Transcribe. See cache, pre-fetch and offline use and CLI and model packages.

Models by role

Tier 1 is ungated MIT/Apache-2.0/BSD and may be redistributed and selected by default. Tier 2 loads only from pinned publisher files, never from our mirror or model packages, never by default. Local conversion stays in your cache. Licence, revision, file digests and attribution travel in ModelInfo. Only the listed licence-cleared Moonshine checkpoints are supported; Community License checkpoints are excluded. See gated models and third-party notices.

Privacy and consent

Speaker embeddings are biometric data. Obtain informed consent, limit retention and access, and check applicable law, including Illinois BIPA and GDPR Article 9, before enrollment, identification or verification. The library does not persist embeddings; applications remain responsible for audio, transcripts, logs and any copies they retain. This is not legal advice.

Guides

• Long files, live events, word timing
• Diarization, speaker recognition, biasing
• Custom CTC/transducer graphs, model cache, tier-2 access
• Transcript JSON schema v1, containers and captions, CLI
• Migration: Whisper.net, sherpa-onnx .NET, Azure Speech SDK, WhisperX concepts

Build and sample checks

Set ONESPEECH_ARTIFACTS to a directory outside the checkout. Run dotnet restore, then dotnet build --no-restore. Run dotnet run --project tests/DocsSamples --no-build -- . to exercise the no-download document examples and check local documentation links. Model-backed examples are compile-checked without downloading weights; run them with your own consented audio and fetched models.