Eleven rows on Sume return MP3 only: use Sonic when you need WAV

The ElevenLabs rows in Sume's TTS Router, eleven-v4, eleven-v4-turbo, eleven-v3, eleven-multilingual-v2 and eleven-turbo-v2.5, return MP3 only. Sonic rows return MP3 or WAV. If a step after speech needs a WAV file, ask a sonic-* id for it. An output_format an Eleven row does not list is refused with a 400 that names the allowed formats.

This is read from the router catalog and request validation in Sume's code on main, 2026-10-11. GET /v1/tts-router/models/{id} returns each row's capabilities.output_formats.

Which output formats does each row accept?

Format ids are container_sampleRate[_kbps]. The default is mp3_44100_128 on every row, and it is the only format on three of the Eleven rows.

What else do the Eleven rows refuse?

The Eleven rows take a transcript, a voice id, a language and a few voice settings. They refuse several fields that the Sonic rows keep from TTS 1.0:

• avatar_id and avatar_handle: an Eleven row needs an ElevenLabs voice id of 20 letters and digits from GET /v1/tts-router/voices?family=eleven, not a Sume avatar.
• timestamps and segmentation: no word timings or per-sentence slices from the job. Run the finished audio through Sume speech-to-text if you need timings.
• transcript_source, generation_config and pronunciation_dict_id: these belong to the Sonic rows.
• voice_settings values outside the range the row lists, for example stability outside 0 to 1.

When does the WAV rule matter?

Sume's TTS 1.0 reference tells callers to request wav with pcm_s16le at 44,100 Hz when the speech feeds an avatar mux, and an avatar-based talking video starts from an avatar's own voice, which an Eleven row cannot select. So the Eleven rows suit narration that goes straight into a timeline as an MP3 voice track, while a talking-avatar clip is a Sonic job.

See Add a voiceover to a video for how a finished voice file joins a video.

How do I confirm the formats before I submit?

List the catalog once, then pick the id. The first call below prints the output formats the eleven-v4 row accepts, and nothing is billed because it reads the catalog, not a job.

The response also carries voice_family, so you know which voice list to call next. Poll jobs as described in Jobs and results.

curl -s https://api.sume.com/v1/tts-router/models/eleven-v4 \
  -H "Authorization: Bearer $SUME_API_KEY" \
  | jq '{id, voice_family, formats: .capabilities.output_formats}'