Alignment

Forced alignment of a known transcript against uploaded WAV bytes (POST /v1/audio/alignments, multipart only). The model must be configured with task: "align", mode: "offline".

Align a transcript

using Mabs.AudioCpp;
using Mabs.AudioCpp.Models;

AlignmentResult aligned = await client.Alignments.RunAsync(new AlignmentRequest
{
    Model = "qwen3-align",
    Text = "hello world",
    Language = "en", // required by the Qwen3 Forced Aligner
    AudioFile = File.ReadAllBytes("hello.wav"),
    AudioFileName = "hello.wav",
});

foreach (AudioCppWord word in aligned.Words)
{
    Console.WriteLine($"{word.Word} {word.Start}-{word.End} (samples {word.StartSample}-{word.EndSample}) conf {word.Confidence}");
}

Words carries seconds (Start/End) plus sample offsets (StartSample/EndSample); Timing is the run timing.

Align and reuse the timestamps

Word timestamps from an alignment feed a caption or karaoke pipeline:

AlignmentResult aligned = await client.Alignments.RunAsync(new AlignmentRequest
{
    Model = "crisperwhisper",
    Text = transcript,
    AudioFile = wavBytes,
    AudioFileName = "take.wav",
});

foreach (AudioCppWord word in aligned.Words)
{
    if (word.StartSample is long start && word.EndSample is long end)
        Console.WriteLine($"{start}\t{end}\t{word.Word}");
}

There is no streaming variant for alignment — the route is multipart, single-shot. See Error handling for failure shapes.