[Enhancement] Add Flux model support for low-latency streaming transcription

[deepgram-robot]

Summary

Add support for the Flux model in the Java SDK's streaming transcription client, enabling developers to use Deepgram's lowest-latency streaming STT model for real-time applications.

Problem it solves

Developers building latency-sensitive Java applications — such as voice agents, real-time captioning systems, and interactive voice response (IVR) platforms — need access to the Flux model's sub-200ms latency. The Flux model uses a different API endpoint (/v2/listen) with distinct connection parameters and response formats compared to Nova-3 streaming. Without SDK support, Java developers must manage raw WebSocket connections manually, losing the type safety and ergonomics the SDK provides.

Proposed API

// Flux streaming transcription
FluxTranscriptionOptions options = FluxTranscriptionOptions.builder()
    .model("flux")
    .encoding("linear16")
    .sampleRate(16000)
    .smartFormat(true)
    .language("en")
    .build();

FluxLiveClient client = deepgramClient.listen()
    .flux()
    .connect(options);

client.onTranscript(transcript -> {
    System.out.println("Transcript: " + transcript.getText());
    System.out.println("Is final: " + transcript.isFinal());
});

client.onUtteranceEnd(event -> {
    System.out.println("Utterance complete");
});

// Send audio data
client.send(audioBytes);

// Finalize and close
client.finalize();
client.close();

Acceptance criteria

• Flux model accessible via listen().flux().connect() or equivalent
• Supports all Flux-specific parameters (encoding, sample_rate, language, keyterms)
• Typed response models for Flux transcript events
• KeepAlive mechanism for long-running connections
• Documented with usage example
• Compatible with existing SDK architecture and event model

---

Raised by the DX intelligence system.