Give your tools a voice — speech to text and back, 25 languages, up to ~19× faster than Whisper. On your machine.
kesha-voice-kit defines 4 tools with reasonable naming (verb_noun pattern: synthesize_speech, transcribe_audio, list_voices, list_languages) and descriptions present for all tools. However, descriptions are brief (50-100 chars) and lack LLM-optimized guidance on when to use each tool vs alternatives. Parameter schemas are present with types and descriptions, but lack constraints (enums, min/max ranges). synthesize_speech accepts optional voice/rate/format but provides no enum for format values in the visible schema. transcribe_audio's path parameter warns about absolute vs relative paths but lacks validation guidance. Error handling is minimal, no recovery guidance or actionable error messages visible in tool definitions. Output schemas are not documented. Tool composition is sound (each does one thing), but descriptions don't explain dependencies or prerequisites.
List languages available for synthesis, derived from installed voices.
List installed TTS voices with their engine and language.
Synthesize speech from text into an audio file and return a resource link. Omit voice to auto-route by language (male defaults en-am_michael / ru-vosk-m02).
Transcribe a local audio file to text. Set timestamps for segment timings.
Descriptions lack LLM-optimized guidance. synthesize_speech (50 chars) and transcribe_audio (60 chars) are below the 100-char minimum for actionable context. No explanation of when to call list_voices vs list_languages, or how voice selection affects output quality.
synthesize_speech format parameter lacks enum constraint in visible schema. Description mentions 'wav, ogg-opus, flac' but no formal enum declaration prevents LLM from selecting invalid values.
No output schemas documented for any tool. LLMs cannot plan downstream calls or extract required fields (e.g., does synthesize_speech return a file path, URL, or resource URI?).
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 63 | 2025-06-18+ | v2 |
synthesize_speech rate parameter lacks min/max bounds. Description says '0.5-2.0' but no JSON Schema constraints (minimum/maximum) prevent LLM from passing invalid values like 10.0.
No error handling guidance visible. Tools lack recovery hints (e.g., 'If transcription fails, try a different audio format' or 'If voice not found, call list_voices first').