Sume TTS 400: send transcript or transcript_source, never both

Sume TTS 1.0 takes exactly one text input: transcript, a literal string, or transcript_source, a reference to a script revision. Sending both, or sending neither, is rejected.

The two shapes

• transcript: literal text, up to 20,000 characters. Fine for generic requests.
• transcript_source: an object with script_revision_id and sentence_ids (up to 1,000). Sume resolves the text server side.
• Live-commerce Formats require transcript_source. A literal transcript there fails with tts_text_source_required.

Check before you send

This Python function runs offline and mirrors the rule, so a bad body fails in your tests instead of at the API.

def check_body(body):
    has_text = "transcript" in body
    has_source = "transcript_source" in body
    if has_text == has_source:
        raise ValueError("send exactly one of transcript or transcript_source")
    if has_text and len(body["transcript"]) > 20000:
        raise ValueError("transcript over 20,000 characters")
    if has_source:
        ids = body["transcript_source"].get("sentence_ids", [])
        if not 1 <= len(ids) <= 1000:
            raise ValueError("sentence_ids must hold 1 to 1000 ids")
    return True


print(check_body({"transcript": "Hello", "language": "en"}))
try:
    check_body({"transcript": "Hello", "transcript_source": {"sentence_ids": ["a"]}})
except ValueError as e:
    print("rejected:", e)

Why it is strict

Source-bound requests are resolved before the credit reservation, so a stale revision or wrong workspace fails closed and charges nothing. Because the text is chosen by id, retries with the same idempotency key reuse the same job instead of drifting.