Kotoba Text-to-Speech (TTS) Streaming

Kotoba TTS v1.9 accepts text input via SageMaker bidirectional streaming (HTTP/2) at the path /invocations-bidirectional-stream. Standard POST /invocations and batch transform jobs are NOT supported.

A synthesis session is a sequence of JSON events sent over the bidirectional stream:

1. open (required, first event)

   • language (string, optional, default ja): ISO-639-1 language code.
   • speaker_id (string, optional, default default): preset voice. Male: male-01 (default and man-1 are the same voice). Female: female-01 (woman-1 is the same voice), female-02, female-03, female-04. man-1 and woman-1 are deprecated and will be removed in a future version.
   • voice_settings (object, optional): the session's default voice controls (speed, pitch, volume; see response.create). Each response.create may override them field by field.
   • format (string, optional, default float32): output audio format. One of float32 (little-endian float32), pcm16 (little-endian signed 16-bit), mulaw (8-bit G.711 mu-law), or opus (Ogg/Opus). The former spellings pcm_f32 and pcm_16 are still accepted.
   • sample_rate (integer, optional, default 24000): output sample rate in Hz. PCM formats (float32, pcm16) accept 8000, 16000, or 24000. mulaw is always emitted at 8000 and opus at 24000, so this field is ignored for them. The negotiated format and sample_rate are echoed on session.created.
   • spk_ref_audio_tokens (optional): speaker reference tokens for the requested voice.

2. response.create (one synthesis turn; repeatable, and may be sent without waiting for the previous response.done — see Pipelining below)

   • text (string, required): the full text to synthesize. The model is one-shot, so the text is not streamed in chunks.
   • response_id (string, optional): client-supplied correlation id; echoed on response.queued / response.created / audio.chunk / response.done. Must be unique among outstanding (active or queued) responses, and reusable once its response.done has been delivered. When omitted the server assigns a UUID.
   • with_timestamps (boolean, optional, default false): when true, audio.chunk events for this response carry alignments records mapping spans of the request text to times in the output audio (see Model output details).
   • voice_settings (object, optional): voice controls for this response. Fields left out keep the session default from open.
     • speed (number, optional, default 1.0): speaking rate from 0.7 to 1.5 without changing the pitch; 1.5 is 1.5x faster.
     • pitch (number, optional, default 0): pitch shift in semitones from -4 to +4 without changing the speaking rate.
     • volume (number, optional, default 1.0): loudness multiplier from 0.0 to 2.0; 0.0 is silent. Above 1.0 the audio is limited at full scale, so the loudest passages may clip slightly.
   • max_length (integer, optional): accepted for compatibility with earlier versions and ignored; the output length follows the text.

3. response.cancel (optional)

   • response_id (string, optional): cancel the response with this id, whether it is the active response or one still queued. Without it, the active response is cancelled.

Pipelining: while a response is active, additional response.create events are accepted with a response.queued event (carrying the same response.id) and synthesized one at a time in FIFO order. The pending queue holds at most 15 responses; exceeding it yields a non-fatal error (too many queued responses).

Each JSON event must be sent as a UTF-8 text frame. With the SageMaker HTTP/2 SDK, construct RequestPayloadPart with data_type="UTF8" so the payload is forwarded as text.