SpeechRecognizer

classin packageklyn.genai.stt

public class SpeechRecognizer implements AutoClosable

SpeechRecognizer

All Implemented Interfaces: AutoClosable

Resident speech-to-text engine, implemented in Klyn/C++ without an external inference library or process. Supports Whisper HF Safetensors and Qwen3-ASR Safetensors/paired GGUF checkpoints; formats never substitute one architecture for another. Projections and encoder attention use CPU/NVIDIA; DSP and decoder attention use synchronized CPU storage. Auto chooses GPU when available; an explicit GPU never silently falls back. NPU is not available yet. Single owner: finish/close its stream before other recognition calls.

Example

try speech = SpeechRecognizer.load("openai/whisper-base"): print(speech.transcribe(AudioClip.load(FilePath("recording.wav")), language="fr").text)

Properties

Modifier and Type	Member	Description
public readonly property	device `device as Device`	Concrete execution device.

Constructors

Modifier and Type	Member	Description
public	SpeechRecognizer `SpeechRecognizer(notnull folder as FolderPath, notnull device as Device = Device.auto, verbose as Boolean = false)`	Opens a local HF snapshot with config/tokenizer/generation_config and Safetensors shards.
public	SpeechRecognizer `SpeechRecognizer(notnull path as FilePath, notnull projector as FilePath, notnull device as Device = Device.auto, verbose as Boolean = false)`	Opens a GGUF speech decoder with its matching audio projector.

Methods

Modifier and Type	Member	Description
public override	close `close() as Void`	Cancels/joins active recognition before releasing model storage.
public static	load `load(notnull repository as String, quantization as ModelQuantization = ModelQuantization.AUTO, notnull device as Device = Device.auto, revision as String = "main", offline as Boolean = false, verbose as Boolean = false, token as String = "", format as ModelFormat = ModelFormat.AUTO) as SpeechRecognizer`	Downloads a verified GGUF pair or original Safetensors checkpoint.
public	stream `stream(notnull microphone as Microphone, notnull language as String = "auto", minimumSpeechLevel as Double = 0.0003, partialResults as Boolean = true) as TranscriptionStream`	Starts one bounded, cancellable live transcription worker.
public	transcribe `transcribe(notnull audio as AudioClip, notnull language as String = "auto") as Transcription`	Transcribes finite PCM, downmixing/resampling to mono 16 kHz.
public	transcribe `transcribe(notnull path as FilePath, notnull language as String = "auto") as Transcription`	Transcribes a WAV file without playback.

Inherited Methods

methodInherited Methods from AutoClosable: close