Text-to-speech with voice cloning. Returns download URL. Supports multiple models: standard (Chatterbox), turbo (fast with paralinguistic tags), and fish (Fish Speech multilingual with emotional markers).
SolSpeak TTS has well-structured schemas with proper types and enums, and descriptions are present for all tools and most parameters. However, several critical issues limit the score: (1) parameter descriptions lack actionable constraints (e.g., no min/max for numeric ranges, no format specs for timestamps); (2) error handling is absent, no guidance on what to do if voice cloning fails or if a model doesn't support a parameter; (3) the delete_voice tool exposes a password parameter, violating secret-injection patterns; (4) some parameter relationships are undocumented (e.g., voice_text is 'REQUIRED for fish model' but this is buried in instructions, not in the parameter description itself); (5) output schemas are not documented, callers don't know what fields the download URL response contains or how to chain results to other tools.
Clone a voice directly from a YouTube video. Include transcript for fish model support.
Delete a saved voice from the voices directory. Requires password.
Generate a multi-voice conversation as a single audio file.
List all saved voices available for voice cloning.
Set or update the transcript for a saved voice. Required for fish model voice cloning.
Generate speech from text using SolSpeak TTS. Returns download URL.
delete_voice exposes password as a tool parameter, violating secret-injection pattern. Passwords must never appear in tool parameters, they leak into agent traces and logs.
No output schemas documented. Callers don't know what fields text_to_speech returns (e.g., is it {url, duration, file_size}? {download_url, expires_at}?). This breaks tool chaining and forces LLMs to guess.
Numeric parameters lack min/max constraints in descriptions. E.g., exaggeration, cfg_weight, temperature, top_p, repetition_penalty have ranges (0.0-1.0, 0.9-2.0) but no validation guidance. LLMs may pass out-of-range values.
Parameter dependencies are undocumented in parameter descriptions. E.g., voice_text is 'REQUIRED for fish model voice cloning' but this constraint is only in the MCP_INSTRUCTIONS, not in the parameter schema. LLMs won't see it.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | D | 58 | 2026-07-28+ | v2 |
No error handling or recovery guidance. If voice cloning fails, if a YouTube URL is invalid, or if a model doesn't support a parameter, there's no documented error response or what the LLM should do next.