Speech MCP Server with command-line interface and Kokoro TTS support. Provides voice interaction with speech recognition and text-to-speech capabilities.
Speech MCP provides 12 tools with inconsistent schema completeness and description quality. Most tools have action-verb naming (listen, speak, transcribe_audio) which is good. However, schemas vary significantly in detail: some parameters lack type information or descriptions. Descriptions range from adequate (speak, listen) to minimal. Several tools are grouped around streaming transcription (start_streaming_transcription, add_streaming_audio_chunk, stop_streaming_transcription, get_current_streaming_transcription), which shows composition issues, three tools manage stateful streaming sessions, requiring users to track session_id across calls. Error handling is minimal; no recovery guidance or validation feedback. No input validation constraints (enums, ranges) are documented. Output schemas are not documented, LLMs cannot predict return structure. This is a domain-specific tool set with moderate quality, but falls short of production standards.
Add an audio chunk to an active streaming transcription session
Get list of available audio input and output devices
Get list of available text-to-speech voices
Get the current partial transcription from an active streaming session
Get the current speech processing state (listening, speaking, idle)
Listen for audio input and transcribe speech to text
Set the audio input device for speech recognition
Set the text-to-speech voice for speech output
Four tools (get_available_voices, get_audio_devices, get_speech_state, and partially others) have NO visible input schemas in the provided code. For get_available_voices and get_audio_devices, empty input {} is shown, but no output schema is documented. LLMs cannot predict what these tools return.
Streaming transcription tools (start_streaming_transcription, add_streaming_audio_chunk, stop_streaming_transcription, get_current_streaming_transcription) form a stateful session-based workflow. The session_id must be tracked across multiple calls. This violates the single-responsibility principle and complicates LLM reasoning. Consider consolidating into a single streaming_transcribe tool or providing a higher-level abstraction.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 52 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 26 | - | v1 |
Convert text to speech and play audio output
Start streaming transcription of continuous audio input
Stop an active streaming transcription session and get final result
Transcribe audio file to text using speech recognition
Descriptions are generic and lack LLM-optimization. Example: 'Get list of available text-to-speech voices' (50 chars) does not explain WHEN to call this tool, what fields are returned, or how to use the results. Descriptions should be 50-200 chars and answer: What does it do? When to use it? What does it return?
No input validation constraints documented. Parameters like 'timeout' (number), 'speed' (0.5-2.0), and 'device_index' lack documented ranges or validation rules. 'voice' parameter accepts free-form string with no enum of valid voices. LLMs will hallucinate invalid values.
No output schemas documented. LLMs cannot plan downstream tool calls or extract return data. Example: transcribe_audio returns 'text' but does not specify if it includes timestamps, speaker labels, or confidence scores, yet input offers include_timestamps and detect_speakers options.
No error handling guidance. Tools fail silently or with generic exceptions. Example: set_tts_voice accepts a voice string, what happens if the voice is invalid? 'Invalid voice' is all an LLM gets. Should return 'Invalid voice: got "xyz". Available voices: [list]' to enable self-correction.
Parameters named 'voice' and 'device_index' are ambiguous. Should be 'voice_id' or 'voice_name' to clarify whether the parameter accepts system IDs or human-readable names. Similarly, 'device_index' is opaque, should clarify it is a numeric index returned by get_audio_devices.
Parameter 'include_timestamps' (boolean) lacks documentation of output format. If true, does the tool return an array of words with timings? A JSON structure? This ambiguity forces LLMs to guess.