inference4j/whisper-tiny-genai

Whisper Tiny — ONNX (onnxruntime-genai)

ONNX export of openai/whisper-tiny in onnxruntime-genai format for autoregressive speech-to-text and translation. Produced with onnxruntime-genai's model builder (FP32, CPU).

Converted for use with inference4j, an inference-only AI library for Java.

Usage with inference4j

Transcription

try (var whisper = WhisperSpeechModel.builder()
        .modelId("inference4j/whisper-tiny-genai")
        .build()) {
    Transcription result = whisper.transcribe(Path.of("audio.wav"));
    System.out.println(result.text());
}

Translation (any language to English)

try (var whisper = WhisperSpeechModel.builder()
        .modelId("inference4j/whisper-tiny-genai")
        .language("fr")
        .task(WhisperTask.TRANSLATE)
        .build()) {
    Transcription result = whisper.transcribe(Path.of("french-audio.wav"));
    System.out.println(result.text());  // English text
}

Model Details

Property	Value
Architecture	Whisper tiny (encoder-decoder, 39M parameters)
Task	Speech-to-text / translation
Languages	99 languages
Audio input	WAV, resampled to 16 kHz mono
Max audio chunk	30 seconds (inference4j splits longer audio)
Precision	FP32
ONNX format	onnxruntime-genai

Original Paper

Radford, A., Kim, J. W., Xu, T., et al. (2022). Robust Speech Recognition via Large-Scale Weak Supervision. arXiv:2212.04356

License

The original Whisper model is released under the MIT License by OpenAI.