Speech to text¶

The SDK supports complete-file transcription and live WebSocket sessions.

Complete file¶

val transcript = elevenLabs.speechToText.transcribe(
    SpeechToTextRequest(
        audio = recordedAudio,
        fileName = "recording.m4a",
        contentType = "audio/mp4",
        modelId = "scribe_v2",
        diarize = true,
    ),
)

println(transcript.text)
transcript.words.forEach { word ->
    println("${word.startSeconds}: ${word.text} (${word.speakerId})")
}

The request retains the input ByteArray to avoid duplicating large media. Do not mutate that array until transcribe() returns.

Realtime transcription¶

Declare the format that your microphone or audio source produces:

val session = elevenLabs.speechToText.openRealtimeSession(
    options = RealtimeSttOptions(
        audioFormat = RealtimeSttAudioFormat.Pcm16000,
        commitStrategy = RealtimeSttCommitStrategy.VoiceActivityDetection,
        includeTimestamps = true,
    ),
)

Collect events while sending audio:

coroutineScope {
    launch {
        session.events.collect { event ->
            when (event) {
                is RealtimeSttEvent.PartialTranscript -> showLiveText(event.text)
                is RealtimeSttEvent.CommittedTranscript -> saveText(event.text)
                else -> Unit
            }
        }
    }

    microphone.pcm16MonoChunks(sampleRate = 16_000).collect { chunk ->
        session.sendAudio(chunk)
    }
    session.commit()
    session.close()
}

With manual commits, call commit() at each desired transcript boundary. With voice activity detection, ElevenLabs determines the boundaries, though an explicit final commit can still be sent.