List Voices

List TTS voices available to the current workspace.

Every filter accepts repeat-key OR semantics: ?gender=female&gender=neutral&category=narration&source=platform&source=workspace. Filters combine across fields with AND; within a field, values OR.

Hybrid search: when search is a non-empty string, semantic search is enabled for the deployment (VOICE_SEMANTIC_SEARCH_ENABLED), and the text embedder is available, search runs a hybrid of SEMANTIC relevance — meaning, not keywords, so “cheerful” also surfaces “bright/upbeat” voices, over the platform voices that carry a profile embedding — AND a keyword arm that matches the voice NAME (plus descriptor/tags), which also surfaces voices with no embedding and workspace-owned voices in scope, so a voice literally named by the query is found. Both arms honor the same filters and are fused into one ranking. In that mode sort/order are IGNORED (relevance order wins). If semantic search is disabled, the embedder is unavailable, or the embed call fails, search transparently falls back to the lexical name/tag/descriptor match described below — the response shape (ApiCountedListResponse[VoiceOut]) is identical either way.

language matches a voice when any of its declared locales matches any requested value. A voice with no declared locales matches NO language filter — it must positively declare a locale to surface under it. This holds for platform and user-uploaded voices alike: an unclassified platform voice (catalog gap) is not treated as general-use, and a user-uploaded/cloned voice with no locale stays “language unknown” pending clone-flow detection.

Passing language also fills language_sample_url on every row that has a clip in it — the same audio GET /voices/{voice_id}/preview?language= serves, so a caller auditioning a shortlist can play the locale-correct take straight from the list instead of a request per voice. language_sample_locale reports the region actually served. Both are null without the filter; sample_url is unaffected and still does not follow it.

Platform voices are restricted to the officially supported locales: a platform voice that declares no official locale is not returned, and the locale arrays on the voices that are returned (supported_languages, model_capabilities[].supported_languages, preview_locales) list only official locales. A bare family code counts as official when the family is supported (ko qualifies because ko-kr is), matching the language filter above. Voices your workspace owns — imported, recorded, or uploaded — are exempt from both the exclusion and the narrowing: they routinely carry no declared locale at all, and hiding them would remove a customer’s own voices from their own list.

Multi-sort: sort and order are parallel lists. ?sort=uses_count&sort=name&order=desc&order=asc orders primarily by uses_count DESC, secondarily by name ASC. When order is shorter than sort, missing entries default per-field: name=asc, created_at=desc, uses_count=desc. When sort is omitted, list defaults to newest-first (or most-recently-favorited-first if favorites_only=true). Every sort path appends Voice.id ASC as a deterministic tiebreaker for pagination stability.