Roles and sessions

transcribe.cpp 0.4 gives a loaded model one or more roles: a kind of work it can do. The roles are ASR (speech to text), DIARIZE (speaker turns), LANGID (language identification) and VAD (voice activity detection). A model serves one or more of them, fixed at load and never changed for its lifetime.

Model.Roles reads them:

if (model.Roles.HasFlag(TranscribeRole.Asr))
{
    using var session = model.CreateSession();
    // session.Run(...)
}

Entry points that need a role the model lacks throw a TranscribeException with Status.ErrUnsupportedRole (20). In particular, Model.GetCapabilities is an ASR query in 0.4: on a model without the ASR role it throws rather than returning a zeroed struct. Ask Model.Roles first.

The four session types

Role	Create with	Run
ASR	Model.CreateSession	Session.Run (offline) / Session.CreateStream
DIARIZE	Model.CreateDiarizeSession	DiarizeSession.Run → speaker segments
LANGID	Model.CreateLangIdSession	LangIdSession.Run → ranked language candidates
VAD	Model.CreateVadSession	VadSession.Run → speech probabilities and segments

Model.GetDiarizeInfo(), GetLangIdInfo() and GetVadInfo() return each role’s own facts (sample rate, and the speaker, label or frame counts), and throw ErrUnsupportedRole for a model without that role.

Diarization

Diarization is offline: upstream has no stream entry point for the DIARIZE role. A Sortformer model carries its latency/quality preset on the diarize params:

using var model = Model.Load("diar_streaming_sortformer_4spk-v2.1-Q8_0.gguf", /* … */);
using var diarize = model.CreateDiarizeSession();
using var preset = new SortformerDiarizeExtBuilder()
    .WithPreset(SortformerPreset.SortformerPresetDefault);
var segments = diarize.Run(pcm, p => p.WithSortformerExt(preset));

The speaker-attributing ASR models (MOSS, granite, multitalker parakeet) are unchanged: they still attribute speakers through the run-time toggle described on Diarization.

Language identification and VAD

LangIdSession.Run ranks the model’s labels (optionally restricted to a set of codes with LangIdParamsBuilder.WithAllowed). The model has a hard cap on the clip (30 s for VoxLingua107), so the CLI and the app identify from the head of the audio and say so.

VadSession.Run returns one speech probability per frame plus the speech segments, and Feed/Flush/Reset drive the streaming form. Silero VAD is chunked-unbounded upstream, so the CLI and the app score the file one window at a time (1 s overlap) and union the intervals across the seams — memory stays one window, not the whole file. The CLI’s --chunk sets that window.

Each window is scored independently, so a seam can bridge a short silence that a one-pass run keeps as a gap (and, less often, the reverse). The speech found is the same and its total duration matches to within a second, but the exact segment boundaries at a seam are not identical to a single pass — that is the cost of not holding the whole file, and it is what upstream’s chunked-unbounded strategy means.

Diarization is the opposite: a hard cap with no streaming entry point, and speaker ids are per run, so a file is not split — windowing it would renumber the speakers at every seam. It is run whole.

Backend allowlist

0.4 lets you restrict which compute backends initialize:

Backends.InitDefault(BackendMask.Cpu | BackendMask.Vulkan);

Backends.AllowedBackends reports the mask in effect.

What is verified

The role enum, the parameter builders and the role/capability contract are covered by UpstreamV04Tests against a real ASR model: a Whisper model reports Asr and refuses the other three roles with UNSUPPORTED_ROLE.

The three sessions are exercised end to end against the real weights, which scripts/run-integration-tests.sh fetches (each pinned to a HuggingFace revision and verified by sha256):

Session	Weights	Test
VAD	Silero VAD v6.2, 1.2 MB	VadSessionTests
LANGID	VoxLingua107 ECAPA-TDNN Q8_0, 27 MB	LangIdSessionTests
DIARIZE	Sortformer Q8_0, 139 MB	DiarizeSessionTests

They cover the role check, the per-role info, a run and its results, the VAD streaming form, timings, disposal, rejection of non-finite audio, and cancellation (both before the call and mid-run). The JFK sample is the input; the assertions are properties (well-formed segments, English identified, the allowed set respected), not recorded golden values that a model change would invalidate.