val config = OfflineRecognizerConfig(
    featConfig = FeatureConfig(sampleRate = 16000, featureDim = 128),
    modelConfig = OfflineModelConfig(
        whisper = OfflineWhisperModelConfig(
            encoder = "$modelDir/large-v3-turbo-encoder.int8.onnx",
            decoder = "$modelDir/large-v3-turbo-decoder.int8.onnx",
            language = "de",
            task = "transcribe"
        ),
        tokens = "$modelDir/large-v3-turbo-tokens.txt",
        numThreads = 4,
        provider = "cpu"
    ),
    decodingMethod = "greedy_search"
)

Why this model?

The multilingual openai/whisper-large-v3-turbo achieves ~14% WER on German benchmarks. This German fine-tune brings that down to ~5% — a significant improvement for German-language applications. The sherpa-onnx build can use the GPU of Mobile devices to option the best posible performance.

MolyProduction/whisper-large-v3-turbo-german-sherpa-onnx

whisper-large-v3-turbo-german — sherpa-onnx INT8

Files

Usage (sherpa-onnx Android)

MolyProduction/
whisper-large-v3-turbo-german-sherpa-onnx