Real-time streaming

StreamSession turns a live audio source into a growing transcript: you feed it chunks as they arrive and read the current text at any point. The wrapper exposes the upstream stream extension for the models that have one.

stream.Begin();
int chunkSize = 16000; // 1 second
for (int i = 0; i < pcm.Length; i += chunkSize)
{
    int length = Math.Min(chunkSize, pcm.Length - i);
    var chunk = pcm.AsSpan(i, length);
    stream.Feed(chunk);
}
stream.Complete();
var text = stream.GetCurrentText();

Begin() resets the stream, Feed() submits one chunk of 16 kHz mono float PCM and returns a StreamUpdateResult, Complete() flushes what is buffered and also returns a StreamUpdateResult, and GetCurrentText() returns a StreamTextResult carrying the current text as full, committed and tentative. Reading the current text before Complete() gives a partial transcript — that is the point of the incremental API, and it is the reason the CLI does not expose streaming (what the CLI does not do).

Two limits worth knowing

• No speaker attribution. transcribe_stream_params has no diarize field in transcribe.cpp v0.2.4, so a stream cannot request it; we do not invent one. The full reasoning is under Not available: streaming.
• Not thread-safe. A StreamSession is a view over a Session and shares its state — see Concurrency.

Ordering and disposal

A stream keeps native state (including a KV cache) for the session’s lifetime. Dispose the StreamSession before its Model, and do not rely on the GC finalizer for cleanup — see Dispose discipline.