Proposal per docs/ADDING_A_MODEL.md. All API observations below were verified against the live API on 2026-07-17 (probe details in PR to follow).
Provider and model id
Speechify, published model id simba-3.2 (POST /v1/audio/speech on api.speechify.ai). The API validates the field strictly: a bogus id 400s with the valid list (simba-english, simba-multilingual, simba-3.0, simba-3.2), so per-model clips are genuine (no Neuphonic-style shared pool).
Voice cloning support (the inclusion criterion)
Speechify has API cloning: multipart POST /v1/voices with name, one sample file, and a consent declaration (docs: https://docs.speechify.ai). Verified live with a throwaway clone (created, synthesized, deleted). Current blocker: a fresh clone lists only simba-english / simba-multilingual in its models; synthesis with a clone against simba-3.2 returns 400. Simba 3.2 is presently limited to eight stock *_32 shared voices (beatrice/dominic/edmund/geffen/harper/hugh/imogen/wyatt; en-US + en-GB). Until Speechify enables cloned voices for 3.2, the same-voice battle methodology cannot run, so this proposal is a staging request, not a registration request.
Public API availability
Public and reachable; the TTFB bench has a surface (chunked HTTP POST /v1/audio/stream; a single probe saw first audio around 0.9 s, which is an observation, not a bench). Note: back-to-back requests rate-limit quickly (429), so generation needs pacing.
Notable
- Pricing: https://speechify.ai/pricing (as of 2026-07-17).
- The stock 3.2 voices are en-US/en-GB only today;
simba-multilingual covers the wider locale set.
- A companion PR stages the verified transport, a pre-registration row (proposed frozen
arenaApiId: simba-3-2), env wiring, and RUNBOOK notes so registration is a maintainer-only pipeline run once the clone gate lifts.
🤖 Generated with Claude Code
Proposal per
docs/ADDING_A_MODEL.md. All API observations below were verified against the live API on 2026-07-17 (probe details in PR to follow).Provider and model id
Speechify, published model id
simba-3.2(POST /v1/audio/speechonapi.speechify.ai). The API validates the field strictly: a bogus id 400s with the valid list (simba-english,simba-multilingual,simba-3.0,simba-3.2), so per-model clips are genuine (no Neuphonic-style shared pool).Voice cloning support (the inclusion criterion)
Speechify has API cloning: multipart
POST /v1/voiceswithname, onesamplefile, and aconsentdeclaration (docs: https://docs.speechify.ai). Verified live with a throwaway clone (created, synthesized, deleted). Current blocker: a fresh clone lists onlysimba-english/simba-multilingualin itsmodels; synthesis with a clone againstsimba-3.2returns 400. Simba 3.2 is presently limited to eight stock*_32shared voices (beatrice/dominic/edmund/geffen/harper/hugh/imogen/wyatt; en-US + en-GB). Until Speechify enables cloned voices for 3.2, the same-voice battle methodology cannot run, so this proposal is a staging request, not a registration request.Public API availability
Public and reachable; the TTFB bench has a surface (chunked HTTP
POST /v1/audio/stream; a single probe saw first audio around 0.9 s, which is an observation, not a bench). Note: back-to-back requests rate-limit quickly (429), so generation needs pacing.Notable
simba-multilingualcovers the wider locale set.arenaApiId: simba-3-2), env wiring, and RUNBOOK notes so registration is a maintainer-only pipeline run once the clone gate lifts.🤖 Generated with Claude Code