Skip to main content

Generation and streaming

Generate one result​

let result = try await tts.generate(
    "Small models can still have a big voice.",
    voice: .luna,
    speed: 1.05
)

print(result.samples.count)
print(result.sampleRate)     // 24000
print(result.duration)
print(result.effectiveSpeed)

Blank input throws KittenTTSError.emptyInput. Speed is clamped to 0.5–2.0 and then combined with the selected model's voice speed prior.

Stream long input​

for try await chunk in tts.generateStreaming(
    articleText,
    voice: .jasper,
    speed: 1.0
) {
    consume(chunk.samples, sampleRate: chunk.sampleRate)
}

The SDK splits input into sentences and yields one complete KittenTTSResult after each local inference. The stream finishes after every sentence or throws the first synthesis error.

Actor behavior​

KittenTTS is an actor. Await its isolated methods and reuse the same instance instead of constructing a model for every sentence. Inference work is detached with user-initiated priority so UI actors are not occupied by the ONNX call.

Long-form playback​

For gap-controlled narration, feed streamed PCM into an audio engine or queue. Calling speak() separately for every sentence is simpler but waits for each full generation and playback cycle.

Last updated on