Know what's being spoken, before you transcribe it.

using LMKit.Model;
using LMKit.Speech;

var model = LM.LoadFromModelID("whisper-large-turbo3");
var stt   = new SpeechToText(model);

var result = await stt.DetectLanguageAsync("clip.mp3");
Console.WriteLine($"{result.LanguageCode} ({result.Confidence:P0})");

var detected = await stt.DetectLanguageAsync("clip.mp3");

// Most languages run great on the speed-optimised turbo variant.
// For CJK and a handful of right-to-left scripts, the slower large
// model is measurably more accurate. Detection picks the variant.
var modelId = detected.LanguageCode switch
{
    "ja" or "zh" or "ko" or "ar" or "he" => "whisper-large3",
    _                                => "whisper-large-turbo3",
};

var tuned = new SpeechToText(LM.LoadFromModelID(modelId));
var transcript = await tuned.TranscribeAsync("clip.mp3");

var detected = await stt.DetectLanguageAsync("clip.mp3");

stt.Mode = detected.LanguageCode == "en"
    ? SpeechToTextMode.Transcription
    : SpeechToTextMode.Translation;       // any language -> English

await stt.TranscribeAsync("clip.mp3");