Pick an ElevenLabs model by language count: 90+, 70+, 32, 29

Match the model to the language list first: ElevenLabs says eleven_v4 covers 90+ languages, eleven_v3 70+, eleven_flash_v2_5 32, eleven_multilingual_v2 29, and eleven_flash_v2 is English only. None of those is a model id on Sume, whose TTS Router catalog is Cartesia Sonic only; Cartesia states 44 languages for Sonic.

If your script is in a less common language, the count decides before price or latency does.

The ElevenLabs list

The models doc gives the language counts below, along with latency and character limits where it states them. The text-to-speech capability page repeats Flash v2.5 at 32 languages with about 75 ms latency and Multilingual v2 at 29 languages, and lists MP3, PCM, mu-law, A-law and Opus as output formats.

Where Sume fits

The Sume TTS Router offers sonic-3.6, sonic-3.5, sonic-3, sonic-latest and sonic-preview. The Cartesia Sonic page states 44 languages, a number that sits between ElevenLabs' Flash v2.5 and v3 counts. The Sume request carries a BCP-47 language code, and the transcript limit is 20,000 characters, which exceeds the ElevenLabs limit for v4, v3 and Multilingual v2 but not for the Flash models.

Language counts are vendor claims about coverage, not quality. A language being listed does not tell you the voice sounds native, so audition it.

A selection routine

Write down the languages you must support at launch and the ones you might add. Remove any model that does not list the launch languages. Among what remains, run the same short script and compare. If the survivors are all ElevenLabs models, you will use ElevenLabs directly and import the audio into Sume for captions and timeline work. If Sonic covers your list, you can keep the whole flow on Sume.

• Check the exact locale you need, not just the language.
• Treat latency figures as streaming claims; Sume TTS is a job API.
• Keep the language code explicit on every Sume request.
• Re-check the lists in a month, since vendors update them often.