Clone a voice from a full recording of a known script, and speak one clip back.

Where POST /voices takes a reference the caller already cut and trusts ASR to say what is in it, this takes the whole enrollment recording and cuts the reference itself: the script is force-aligned to the audio, every run of whole sentences between 4 and 13 seconds is scored, the best is cut, and its punctuation is rewritten to the pauses the speaker actually left. See api.services.enrollment.

Args: voice_name: The desired name for the cloned voice. voice_tags: Comma-separated tags for the voice. lang_code: Language of the voice. script: The text the speaker read. Defaults to the built-in script for lang_code, which is what GET /voices/assisted/script returns. preview_text: Line to speak back with the new voice. Defaults to a fixed sentence; pass an empty string to skip the preview. fast: Use our own enhancer (~10 s for a typical take) instead of the external one (typically 20-30 s, up to ~40 s). The external one handles clipping and uneven levels better, so clipped takes are likelier to be rejected with fast. Without fast our enhancer runs alongside the external one, and its clone is returned when the external one is unavailable or its side rejects the recording. reference.enhancer says which ran. candidates: Offer several voices to choose between instead of one (see Returns). The recording is enhanced once by each enhancer either way. voice_file: The enrollment recording, the whole script in one take, 15-60 seconds.

Returns: dict: the new voice_id, a reference report describing what was chosen and why, and a preview clip unless one was not requested or could not be produced. The voice is stored with assisted true and the report in assisted_report.

    With `candidates`: a candidate set, a voice per usable pair of consecutive
    sentences (up to MAX_SEGMENTS, the best if there are more, or the single clone's
    passage if none is usable), cut from the external enhancer's output and named
    `<voice_name>_slow_<first sentence>`, plus, without `fast`, the best pair cut from our
    own as `<voice_name>_fast_<n>`; when the external enhancer gives none, the pairs come
    from our own. Every voice is tagged `candidates:<set id>`. `candidates` lists them
    best first, each with its `rank`, `candidate` id, `reference` report and `preview`;
    each enhancer's best also comes as a `top_k` 10 variant (`<id>_top_k_10`, the same
    voice, the same rank), whose preview is that voice synthesised with exactly its
    `top_k` and `temperature` (null leaves the default). Every voice is kept, with
    `assisted` true, `selected` false and its offer, rank, sampling variants and
    reference report in `assisted_report`; POST /voices/assisted/choose records which
    one the user picked.