Answers application/x-ndjson by default: one JSON object per line, written the moment that item finishes, each carrying its audio as base64. A final {"object":"batch.done"} line closes the stream. Send stream: false for one buffered application/json envelope instead.

Each item’s audio is a COMPLETE file — real sizes in its header, never the streaming placeholder — so you can write the decoded bytes straight to disk. Set response_format on the batch, or per item to mix formats in one request; each item carries format, mimeType, sampleRate, channels and bitsPerSample next to its audio.

Audio defaults to MPEG-2 Layer III (MP3), 24 kHz, mono, 128 kbps, with no Xing/LAME header and no ID3 tag — the body starts at the first audio frame. Set response_format for anything else: mp3, mp3_64, mp3_160, wav, wav24, wav32f, flac, flac24, pcm, pcm_f32, aiff, caf, au, mulaw, alaw. The exact numbers for a given response are echoed in the X-Ditto-Audio-* headers; decode and resample from those rather than assuming a rate. GET /v1/audio/format describes every encoding in full, without authentication.

Hi-fidelity formats bill at 1.2× the normal rate, multiplied by the model’s own weighting — the audio is identical, but the bytes are several times larger. mp3 (the default), mp3_64, mulaw and alaw are unchanged at 1×. Each format’s exact multiplier is in GET /v1/audio/format and in the X-Ditto-Audio-Cost-Multiplier response header.

Results are not in request order. Items are reordered server-side so that work sharing a (model, voice, language) runs contiguously — the serving node holds ONE voice slot, so an interleaved order forces a speaker reload on every item, which is the cost this endpoint exists to remove. Streamed lines therefore arrive as they complete; every line carries index (its position in your items array), and the buffered envelope is sorted back into request order for you.

Failure is PER ITEM: a bad voice on one item produces a failed result for that item with the same status/code the single-shot route would have returned, and does not affect the others. The HTTP status is 200 whenever the batch itself was accepted.

Aborting the request stops dispatching further items — in-flight synthesis is cancelled at the node, so cancelling saves GPU.

Up to 64 items and 200000 characters per request. For more, or for work too long to hold a connection open, use POST /v1/audio/batches.

{
  "model": "kokoro",
  "voice": "kokoro-af-heart",
  "language": "en",
  "items": [
    {
      "id": "para-17"
    }
  ]
}