Darkspyre.Inference.Tensor.STT.Whisper 11.19.3-rc.2

Darkspyre.Inference.Tensor.STT.Whisper

Speech transcription with Whisper models inside the application's process, through the public Whisper.net runtime. The model stays loaded across requests on CPU or CUDA. The application owns model storage and the engine's lifetime.

Install

dotnet add package Darkspyre.Inference.Tensor.STT.Whisper

Whisper.net, its CPU and CUDA 12 Linux runtimes and NAudio.Core arrive as NuGet dependencies.

Register

Construct WhisperSpeechToTextInference over a TensorModelPackManager, or call AddTensorEngines() from Darkspyre.Inference.Tensor.Registry, which registers WhisperEngineProvider. TensorSpeechModels.Whisper lists the twelve built-in packs: tiny, base, small and medium with their English-only variants, large-v1, large-v2, large-v3 and large-v3-turbo.

Example

using var packs = new TensorModelPackManager(new TensorModelCatalog(TensorSpeechModels.Whisper), packDirectory);
await using var whisper = new WhisperSpeechToTextInference(packs);
await whisper.StartAsync();
await whisper.LoadAsync(modelKey); // Loaded or Failed arrives through EngineStatusChanged.

SpeechTranscriptionResult result = await whisper.TranscribeAsync(new SpeechTranscriptionRequest(audio));

audio is a SpeechAudio of interleaved float PCM with its sample rate and channel count. TranscribeStreamAsync transcribes a sequence of audio windows and returns one result per window.

Settings

WhisperSetup.CreateSettingsManager opens the engine's settings: the twelve built-in models with their install state, "Runs on", and the download, check, load and backend-check actions.

Limits

• Audio from 8 to 192 kHz with 1 to 8 channels is converted to mono 16 kHz. An empty clip gives an empty result.
• A language code and translation to English are request options. Translation is refused on English-only and turbo models.
• Automatic backend selection prefers CUDA 12 and falls back to CPU. TensorBackendKind.Cpu turns the GPU off; the constructor's gpuDevice picks the GPU.
• Standard F32 and F16 Whisper GGML files load. Quantized variants are not offered.
• Window transcription does not merge overlapping windows, separate speakers or emit partial results.
• Native processing runs in the application's process, and teardown has no fixed deadline.

The Darkspyre.Inference packages

This package is one library of the Darkspyre.Inference family. The family runs AI inference from .NET applications through typed, provider-neutral contracts.

• Darkspyre.Inference.Common holds the contracts every other package shares: requests, results, capabilities, failures and the provider setup model.
• Darkspyre.Inference.Connect.Common holds the conversation and embedding services. Each Connect provider package adds one provider, and AddLlmConnect() discovers every provider package the application references.
• The Tensor packages load and run models inside the application's own process.
• The UI.Settings packages provide the setup editors a host places in its own window or page.

Every package of the family carries the same version. Reference all of them at the same version: each package depends on the exact version of the family packages it uses.

The packages are licensed under the MIT license. LICENSE.md in each package holds the license text.