deepgram-dotnet-conversational-stt

Using Deepgram Conversational STT / Flux (.NET SDK)

This repo does not currently expose a dedicated Flux / conversational STT API surface comparable to the Python SDK's listen.v2.connect(...) + TurnInfo flow.

Use a different skill when:

• You only need standard streaming transcription without turn awareness → deepgram-dotnet-speech-to-text.
• You need a full voice assistant (STT + LLM + TTS) → deepgram-dotnet-voice-agent.

Current repo status

What exists:

• ClientFactory.CreateListenWebSocketClient() returns the latest WebSocket listen client.
• Request model: Deepgram.Models.Listen.v2.WebSocket.LiveSchema.
• Event models: OpenResponse, MetadataResponse, ResultResponse, SpeechStartedResponse, UtteranceEndResponse, CloseResponse, ErrorResponse, UnhandledResponse.
• Control helpers: SendKeepAlive(), SendFinalize(), SendClose(), Send(...).

What is not present in the current repo search:

• No flux model constants or examples.
• No TurnInfo / turn-aware event models.
• No language_hint, eager_eot_threshold, eot_threshold, or similar Flux request properties.
• No README/examples that mention conversational STT explicitly.

Closest supported code path today

using Deepgram;
using Deepgram.Models.Listen.v2.WebSocket;

Library.Initialize(); // reads DEEPGRAM_API_KEY env var
var liveClient = ClientFactory.CreateListenWebSocketClient();

await liveClient.Subscribe(new EventHandler<ResultResponse>((sender, e) =>
{
    var transcript = e.Channel.Alternatives[0].Transcript;
    if (!string.IsNullOrWhiteSpace(transcript))
    {
        Console.WriteLine(transcript);
    }
}));

await liveClient.Subscribe(new EventHandler<UtteranceEndResponse>((sender, e) =>
{
    Console.WriteLine(e.Type);
}));

await liveClient.Connect(new LiveSchema()
{
    Model = "nova-3",
    Encoding = "linear16",
    SampleRate = 16000,
    InterimResults = true,
    UtteranceEnd = "1000",
    VadEvents = true,
});

Treat this as standard live STT, not true Flux parity.

Key params currently available

On LiveSchema: Model, Encoding, SampleRate, InterimResults, UtteranceEnd, VadEvents, Endpointing, NoDelay, Punctuate, SmartFormat, Keywords, Keyterm, Diarize, Redact.

Workflow: adding Flux support to the SDK

If the task requires real Flux parity, follow these steps in order:

1. Add request params — extend Deepgram/Models/Listen/v2/WebSocket/LiveSchema.cs with LanguageHint, EagerEotThreshold, EotThreshold, and other Flux-specific fields. Validate against the AsyncAPI spec.
2. Add response models — create TurnInfo and any turn-aware event types under Deepgram/Models/Listen/v2/WebSocket/. Verify field names match the AsyncAPI spec.
3. Wire events in the client — update Deepgram/Clients/Listen/v2/WebSocket/Client.cs to deserialize and dispatch new event types.
4. Write tests — add unit tests covering serialization of new request params and deserialization of new response types.
5. Add an example — create examples/speech-to-text/websocket/flux/Program.cs demonstrating a Flux session with turn-taking.

Gotchas

1. Flux is not first-class here yet. Do not invent TurnInfo-style .NET models or ConnectFluxAsync(...) helpers that are not backed by real implementation.
2. Listen.v2.WebSocket naming is misleading for Python-parity expectations. It is the newest streaming client, but not a full conversational surface.
3. DeepgramWsClientOptions defaults APIVersion to v1. Inspect connection URIs before assuming /v2/listen behavior.

Example files in this repo

• examples/speech-to-text/websocket/file/Program.cs
• examples/speech-to-text/websocket/http/Program.cs
• examples/speech-to-text/websocket/microphone/Program.cs

References

• In-repo: Deepgram/Clients/Listen/v2/WebSocket/Client.cs, Deepgram/Models/Listen/v2/WebSocket/*.cs
• AsyncAPI (target spec): https://developers.deepgram.com/asyncapi.yaml
• Product docs: https://developers.deepgram.com/reference/speech-to-text/listen-flux