Single tool 'aivis-speech' has a well-structured JSON Schema input with proper type definitions and descriptions for all parameters. However, the tool description is in Japanese, which limits accessibility for English-speaking LLM contexts. More critically, there is no documented output schema, the LLM cannot know what fields to expect in the response. The tool is language/locale-specific (Japanese TTS) and lacks error handling guidance, recovery instructions, or examples of what successful responses look like. Parameter descriptions are present and clear (e.g., 'wait_ms' has min/max bounds), but the tool lacks the comprehensive error classification and recovery patterns expected of production-grade tools.

• Add a clear English description (100 - 150 chars) explaining when and why to use aivis-speech. Example: 'Synthesizes and plays text as speech using Aivis voice model. Use when the agent needs to generate audio output or speak a response. Supports custom voice model, playback synchronization, and speech parameters (rate, pitch, volume).'
• Document the output schema. Specify what fields the response contains (e.g., { 'audio_url': string, 'duration_ms': number, 'model_used': string }) so the LLM knows what data is available.
• Add error handling and recovery guidance to the description. Example: 'Returns error if text is empty, API key invalid, or model UUID not found. If model not found, call list_available_models first. If playback fails, check that audio output is available.'
• Rename tool from 'aivis-speech' to 'synthesize_and_play_speech' or split into two tools: 'synthesize_speech' (returns audio bytes/URL) and 'play_speech' (plays pre-synthesized audio). This follows verb_noun pattern and enables composition.
• For 'model_uuid' parameter, add an optional 'model_name' variant or document how to discover valid UUIDs. Include a reference: 'Use the list_available_models tool to find model UUID if unknown.'
• Clarify 'sync' parameter behavior: does it block the response until playback completes? Add timeout guidance (max 60s?). Specify what 'true' vs 'false' returns. Example: 'sync=true: waits for playback to complete before returning (max 60s); sync=false: returns immediately with audio URL/ID for asynchronous playback.'