fix(speech-to-text): send keyterms as repeated multipart form fields

[fern-api]

Summary

Stopgap fix so speech_to_text.convert(keyterms=[...]) works without upgrading the Python SDK generator (this repo is on generator 4.64.1).

keyterms (an Optional[List[str]] multipart body property) was serialized as a single JSON-encoded form field:

"keyterms": json.dumps(jsonable_encoder(keyterms)),   # -> keyterms=["hello","world"]  (one field)

The API validates each keyterm individually, so the whole JSON-array string is treated as one keyword and rejected (400 invalid_keyword / invalid_keyword_length; [ ] " are also unsupported characters). This is the elevenlabs-python#819 regression.

Change

Pass the list straight through jsonable_encoder so httpx emits repeated form fields (keyterms=hello, keyterms=world), matching pre-2.59.0 behavior and what the API expects:

"keyterms": jsonable_encoder(keyterms) if keyterms is not OMIT else OMIT,

This is byte-for-byte what Fern's Python generator >=5.22.2 now emits (upstream fix: fern-api/fern#17226), so the manual edit and the eventual generator upgrade converge.

Files changed (2)

• src/elevenlabs/speech_to_text/raw_client.py — sync + async convert (2 lines).
• .fernignore — add raw_client.py so SDK regeneration won't overwrite this patch while staying on generator 4.64.1.

Notes

• Scope is limited to keyterms; other fields (entity_detection, additional_formats, etc.) are unchanged.
• Once this repo bumps its Python generator to >=5.22.2, this patch becomes redundant and the .fernignore entry can be removed.

---

Note

Low Risk
Narrow request-encoding fix for an optional STT parameter; no auth, billing, or broader API surface changes.

Overview
Fixes speech_to_text.convert(keyterms=[...]) failing with API 400s (invalid_keyword / invalid_keyword_length) because keyterms was sent as one JSON-encoded multipart field instead of repeated keyterms= fields per term.

In src/elevenlabs/speech_to_text/raw_client.py, sync and async convert now pass jsonable_encoder(keyterms) (or omit when unset) instead of json.dumps(jsonable_encoder(keyterms)), so httpx emits the repeated form fields the API validates individually. Other multipart fields are unchanged.

.fernignore now includes raw_client.py so Fern regeneration on generator 4.64.1 does not revert this stopgap until the SDK bumps to generator ≥5.22.2.

^{Reviewed by Cursor Bugbot for commit 4220b3c. Bugbot is set up for automated code reviews on this repo. Configure here.}

[kraenhansen]

Superseded by #825 (since the regression is not contained to the speech_to_text endpoint)