WebSocket streaming

Use stream_input() when text arrives in fragments, such as output from a language model.

1. Open a session

stream = client.audio.speech.stream_input(
    voice="default",
    response_format="mp3",
    quality="fast",
)

2. Send text and read audio

with stream:
    stream.send_text("The first fragment ")
    stream.send_text("and the final fragment.", flush=True)
    stream.finish()

    for event in stream:
        if event.audio:
            player.write(event.audio)
        if event.is_final:
            break

flush() completes the current utterance. finish() ends input and lets the remaining audio drain. Async sessions use awaited methods and async iteration.

Session options

Option	Accepted values
`response_format`	`pcm` or `mp3`
`output_sample_rate`	8000–48000 Hz
`quality`	`fast` or `expressive`
`seed`	Optional integer
`chunk_length_schedule`	Optional sequence of chunk targets
`auto_mode`	`True` or `False`

Alignment collections are wire-compatible placeholders, not timestamp data.