ASR4J

ASR4J provides various Automatic Speech Recognition (ASR) interface implementations to various ASR providers.

Supported providers:

ASR Provider	Description
Whisper.cpp	Uses native whisper.cpp to run inference locally
vLLM transcription endpoint	Uses a http client to send the PCM data to the vLLM server
vLLM realtime endpoint	Uses a websocket client to process PCM data for the vLLM realtime endpoint

• Whisper.cpp
• vLLM transcription endpoint
• vLLM realtime endpoint (WIP)

Limitations

Currently only AMD64 Linux is supported. Support for other platforms is not planned.

Usage

<dependency>
  <groupId>io.metaloom.asr4j</groupId>
  <artifactId>asr4j</artifactId>
  <version>0.0.1-SNAPSHOT</version>
</dependency>

License

The code of this project is "Apache License" but the license of the models may be different.

Examples

ASR Whisper.cpp

CountDownLatch latch = new CountDownLatch(1);

String MOVIE = "media/jfk.webm";
String LANG = "en";
String MODEL_PATH = "models/ggml-large-v3-turbo.bin";

Whisper whisper = Whisper.create(MODEL_PATH);
whisper.transcribe(MOVIE, LANG).subscribe(new Flow.Subscriber<>() {

	@Override
	public void onSubscribe(Flow.Subscription subscription) {
		subscription.request(Long.MAX_VALUE);
	}

	@Override
	public void onNext(WhisperSegment segment) {
		System.out.println("start: " + segment.getStart());
		System.out.println("end: " + segment.getEnd());
		System.out.println("text: " + segment.getSentence());
	}

	@Override
	public void onError(Throwable throwable) {
		throwable.printStackTrace();
		latch.countDown();

	}

	@Override
	public void onComplete() {
		System.out.println("Transcription complete!");
		latch.countDown();

	}
});
latch.await();

ASR vLLM - Transcribe Client

String MOVIE = "media/jfk.webm";

ASRClient client = ASRClient.newBuilder()
	.setModel(ASRClient.DEFAULT_WHISPER_MODEL_NAME)
	.setBaseURL("http://localhost:8000/v1").build();

HttpResponse<JsonObject> response = client.transcribe(MOVIE);
System.out.println(response.statusCode());
System.out.println(response.body().encodePrettily());

ASR vLLM - Realtime Client

Please note that not all ASR models support this endpoint.

Voxtral Mini 4B supports this endpoint

. venv/bin/activate
export CUDA_VISIBLE_DEVICES=0,1
export VLLM_DISABLE_COMPILE_CACHE=1
vllm serve mistralai/Voxtral-Mini-4B-Realtime-2602 --compilation_config '{"cudagraph_mode": "PIECEWISE"}

String MOVIE = "media/jfk.webm";

ASRClient client = ASRClient.newBuilder()
	.setModel(ASRClient.DEFAULT_VOXTRAL_MODEL_NAME)
	.setBaseURL("http://localhost:8000/v1").build();

client.realtime(MOVIE);

Build

Requirements:

• JDK 25 or newer
• Maven

Releasing

# Set release version and commit changes
mvn versions:set -DgenerateBackupPoms=false
git add pom.xml ; git commit -m "Prepare release"

# Invoke release
mvn clean deploy -Drelease