One TTS job per sentence, then a gapless concat: the Sume recipe

Make narration one sentence per job, then join the takes with Sume timeline audio. The MCP docs use tts_create per sentence as their example of a loop worth running in a script. The join is sample-domain, with no re-TTS and no silence at the seams, so a line you redo later does not force the rest of the narration to be regenerated.

The flow

Why per sentence

A single long job gives you one take. If sentence seven is wrong, you redo the whole job. Per-sentence jobs let you regenerate only line seven and join again. The concat has a 20-part limit, so a script over 20 sentences needs two joins, then a final join of the two outputs: 2 + 1 = 3 jobs, or $0.03 for the joins.

Sume also gives free read tools for scripts that must match an accepted text: tts_source_get returns the accepted-script manifest for tts_create with transcript_source, and tts_source_verify_spine compares the selected TTS jobs with the accepted script.

Keeping the voice the same

A completed text-to-speech job records model_id, voice ({ "mode": "id", "id": "..." }), language, output_format, generation_config and speed. Each is null if the request did not send it. Read these values from the first job and reuse them for later lines so the voice stays consistent across takes.

Channel layout

All parts you concat must share a channel layout, or the worker returns audio_parts_channel_mismatch. Generate every line with the same output settings, and keep wav through the join as the docs advise. Microsoft's announcement lists MAI-Voice-2.1 at $22 per 1M characters (read 2026-10-08); Sume does not list that model, so the flow above applies to the TTS Sume does list.