MCP Server Pronunciation

What JuhongPark/mcp-server-pronunciation MCP server does

JuhongPark/mcp-server-pronunciation MCP server turns microphone input into English-learning feedback for an MCP assistant. The converse tool is intended for open-ended voice chat: it records until silence or the duration limit, transcribes the speech, and returns brief pronunciation, grammar, and fluency notes along with guidance for the assistant. The practice tool is intended for reading a supplied sentence and produces a more detailed pronunciation assessment.

For targeted practice, suggest_sentence selects a sentence by phoneme focus and difficulty, while quick_practice selects one and immediately records the learner. retry repeats the most recent practice target. The available focus values are th, f_v, r_l, vowels, and general; difficulty values are beginner, intermediate, and advanced.

How it works

Audio is captured from the local microphone and processed on the user’s machine. The default transcription model is faster-whisper, and the README states that recorded audio does not leave the machine. Recordings stop automatically after detected silence, although each recording tool also accepts a maximum duration.

When a reference sentence is supplied, the assessment aligns the spoken words with the expected text and can report expected versus produced IPA, weak phonemes, prosody observations, and learner-profile hints. The bundled profile includes Korean-L1 patterns such as r/l confusion, th-to-s substitutions, final cluster deletion, and an intrusive onset vowel. With the optional [phoneme] installation, wav2vec2 CTC forced alignment provides an additional acoustic check intended to reduce transcription-related errors for unusual names and domain terms. Without a reference sentence, assess reports the transcript and prosody information rather than a full word-level pronunciation comparison.

Setup and configuration

The project requires Python 3.11 or newer and a working microphone. The documented launcher is uvx mcp-server-pronunciation; it can also be installed with pip or as a uv tool. Linux users may need to install PortAudio through their system package manager. The default Whisper model uses about 150 MB of disk space, and the optional phoneme extra requires additional model storage.

Run uvx mcp-server-pronunciation doctor to check the environment before connecting an MCP client. pull-model base.en can pre-download the default model. MCP client examples are provided for Claude Desktop, Cursor, Codex CLI, Claude Code, and VS Code. The JuhongPark/mcp-server-pronunciation MCP server runs over stdio in these configurations.

Tools and capabilities

• converse: record, transcribe, and provide quick conversational feedback.
• practice: assess reading of a specified sentence.
• quick_practice: choose a sentence and start a drill.
• retry: record the previous practice sentence again.
• suggest_sentence: return a sentence with its focus and difficulty.
• record: save an unassessed microphone recording as a WAV file.
• assess: analyze the latest recording or a specified WAV file.
• check_mic: list available input devices and identify the default microphone.

Assessment results include rendered reports and structured values such as transcript, clarity percentage, speaking rate, top issue, next action, and retry comparison. The project also supports macOS, Linux, Windows, and WSL2 according to the README.

Limitations and notes

Feedback is a coaching signal, not a standardized-test, clinical, employment, or other high-stakes assessment. The README warns that transcripts, pronunciation results, recording behavior, and runtime stability may be inaccurate or platform-dependent. A microphone is required for recording features, and the optional forced-alignment functionality adds dependencies and model storage. The JuhongPark/mcp-server-pronunciation MCP server currently documents a Korean-L1 learner profile; additional learner profiles are not described as built in.