mcp-server-pronunciation

# mcp-server-pronunciation [](https://glama.ai/mcp/servers/JuhongPark/mcp-server-pronunciation) <!-- mcp-name: io.github.JuhongPark/pronunciation --> > **Accuracy and safety notice** > > This project is a local language-learning practice tool. It may contain bugs, > runtime errors, inaccurate transcripts, inaccurate pronunciation feedback, or > platform-specific recording issues. Pronunciation feedback is a coaching > signal, not a standardized-test, clinical, employment, or high-stakes > assessment. Review outputs carefully before relying on them. See > [DISCLAIMER.md](DISCLAIMER.md). An MCP (Model Context Protocol) server that lets you **talk to your MCP assistant by voice while getting English pronunciation, grammar, and fluency feedback** in the same turn. Use it for casual voice chat with light coaching, or switch to drill mode when you want to practice a specific sentence. Built for Codex CLI, Claude Desktop, Claude Code, Cursor, VS Code, and other MCP clients. Everything runs locally — audio is captured with your mic, transcribed by [`faster-whisper`](https://github.com/SYSTRAN/faster-whisper) on-device, and never leaves your machine. `mcp-name: io.github.JuhongPark/pronunciation` ## Why Voice MCP servers today treat speech as a typing replacement. English tutor MCP servers are text-only. This one combines the two: you speak freely, your assistant replies, and feedback on what you just said (pronunciation, grammar, fluency) surfaces inside the same tool call so the assistant can weave it into a natural reply — or stay out of the way when you're just chatting. ## Features - **Voice conversation** with your MCP assistant. Speak, auto-stop on silence, then let the assistant read your transcript and respond. - **Phoneme-level drill feedback** (when a reference sentence is given): Needleman-Wunsch word alignment, per-word expected vs produced IPA, learner-profile hints, minimal-pair drills, and prosody checks (word stress, final-rise intonation, intra-clause pauses). - **Extensible learner-profile support**: the current rule pack includes Korean-L1 pronunciation-pattern hints and Korean-language tips. Contributions for additional L1 profiles are welcome. - **Whisper-bias mitigation** via optional `[phoneme]` extra: wav2vec2 CTC forced alignment verifies whether the user actually produced each reference word, so rare proper nouns and domain-specific terms that Whisper rewrites toward more common alternatives no longer surface as mispronunciations. - **Inline English feedback in conversation**: pronunciation, grammar (common irregular-verb errors), and fluency (pace + long pauses). - **Drill mode** (`practice`, `quick_practice`, `retry`) for focused sentence practice. - **Local-only**: Whisper model runs on your machine, audio never leaves it. - **Cross-platform**: macOS, Linux, Windows, and WSL2 (recording auto-routes through Windows). - **Fast startup**: lazy imports + background model pre-load keep the MCP handshake under a second. ## Requirements - Python 3.11+ - A working microphone - ~150 MB disk space for the default Whisper model (`base.en`) - Additional ~360 MB if you install the optional `[phoneme]` extra (wav2vec2 weights for forced alignment) - MCP spec: targets `2025-06-18` via the official Python SDK (`mcp>=1.2`) ## Installation ### Stable release Install the latest stable release: ```bash uvx mcp-server-pronunciation ``` For pip users: ```bash pip install mcp-server-pronunciation ``` To pin this release explicitly: ```bash uvx mcp-server-pronunciation@0.3.0 ``` Run `doctor` before relying on the server in a live session: ```bash mcp-server-pronunciation doctor ``` ### General install commands ```bash # Recommended: uvx (no global install, cached between runs) uvx mcp-server-pronunciation # Or install as a uv tool uv tool install mcp-server-pronunciation # Or pip pip install mcp-server-pronunciation # Optional: forced-alignment upgrade for Whisper-bias mitigation + tighter # phoneme-level feedback. Adds ~200 MB of torch CPU wheels. pip install 'mcp-server-pronunciation[phoneme]' ``` ### Linux: install PortAudio first `sounddevice` ships PortAudio inside the wheel on macOS and Windows, but on Linux you need the system library: ```bash # Debian / Ubuntu sudo apt-get install libportaudio2 # Fedora / RHEL sudo dnf install portaudio # Arch sudo pacman -S portaudio # PipeWire-only systems may also need sudo apt-get install pipewire-alsa ``` ### First-time check Before wiring the server into an MCP client, run the preflight: ```bash uvx mcp-server-pronunciation doctor ``` Optional — pre-download the Whisper model (~150 MB) so the first call is instant: ```bash uvx mcp-server-pronunciation pull-model base.en ``` ## Add to your MCP client ### Codex CLI ```bash codex mcp add pronunciation -- uvx mcp-server-pronunciation ``` ### Claude Code ```bash claude mcp add pronunciation -- uvx mcp-server-pronunciation ``` ### Claude Desktop Edit `claude_desktop_config.json`: ```json { "mcpServers": { "pronunciation": { "command": "uvx", "args": ["mcp-server-pronunciation"] } } } ``` On macOS, if Claude Desktop can't find `uvx` (`spawn uvx ENOENT`), use an absolute path. Find it with `which uvx` in your terminal. ### Cursor Add to `~/.cursor/mcp.json`: ```json { "mcpServers": { "pronunciation": { "command": "uvx", "args": ["mcp-server-pronunciation"] } } } ``` ### VS Code (with MCP support) Add to `.vscode/mcp.json` or your user settings: ```json { "servers": { "pronunciation": { "type": "stdio", "command": "uvx", "args": ["mcp-server-pronunciation"] } } } ``` ## Usage Examples ### 1. Voice chat with feedback > **You**: "Let's have a voice chat. I'll ask you about the weekend. Use the converse tool." > > **Assistant** (calls `converse`): *records your speech, transcribes it, notes that you said "buyed" instead of "bought"* > > **Assistant**: "Oh nice — what kind of apples did you **buy**? And by the way, the past tense of 'buy' is 'bought' — small thing, but I noticed it." ### 2. Drill a specific sentence > **You**: "Give me a sentence to practice with 'th' sounds." > > **Assistant** (calls `suggest_sentence` with `focus=th`): "Try this: *The three brothers thought thoroughly about their future.*" > > **You**: "Record me reading it." > > **Assistant** (calls `practice` with that reference): *returns an alignment table (match / sub / ins / del) with per-word acoustic confidence when the `[phoneme]` extra is installed, phoneme-level issues with expected vs produced IPA, learner-profile hints when applicable, minimal-pair drills, and prosody notes (word stress, final-rise intonation, intra-clause pauses).* ### 3. Retry after feedback > **You**: "Let me try again." > > **Assistant** (calls `retry`): *re-records the same target sentence and compares* ## Tools | Tool | Purpose | |---|---| | **`converse`** | **Primary**. Record + transcribe + quick feedback + assistant guidance for natural voice-chat-with-coaching. | | `practice` | Drill mode: record user reading a specific reference sentence, return detailed assessment. | | `quick_practice` | Pick a random sentence (by phoneme focus + difficulty) and drill it. | | `retry` | Re-record the last sentence and compare the new attempt against the previous one. | | `open_voice_panel` | Open the MCP Apps voice panel when the client supports embedded UI. | | `analyze_uploaded_audio` | Analyze WAV audio uploaded by the voice panel and store it as the latest voice capture. | | `start_voice_capture` | Start recording in the background and return a session id immediately. | | `voice_capture_status` | Check whether a background capture is recording, analyzing, done, cancelled, or failed. | | `wait_for_voice_capture` | Wait for a background capture to finish and return transcript + feedback. | | `latest_voice_capture` | Return the most recent background voice capture result. | | `cancel_voice_capture` | Mark a background capture as cancelled before analysis starts. | | `suggest_sentence` | Return a practice sentence without recording. | | `record` | Record audio and save a WAV file (raw, no analysis). | | `assess` | Assess the last recording (or a specified WAV) without re-recording. When given a reference, runs the full drill pipeline (alignment, phoneme diff, learner-profile hints, prosody). | | `check_mic` | List available audio input devices. | Tools that assess speech also return structured MCP output with `transcript`, `clarity_pct`, `speaking_rate_wpm`, `top_issue`, `next_action`, `retry_comparison`, the full machine-readable `assessment`, and the rendered `report_markdown`. MCP clients can use the structured result to offer a retry, surface the top issue, or build a richer practice UI without parsing Markdown. ### Visible voice-capture workflow For MCP clients without an embedded voice UI, use the background capture tools to keep the user informed: ```text start_voice_capture(duration=8, mode="conversation") voice_capture_status(session_id) wait_for_voice_capture(session_id, timeout=30) latest_voice_capture() ``` The status response includes `recording`, `analyzing`, `done`, `error`, or `cancelled`, plus elapsed time, transcript, clarity, speaking rate, feedback markdown, and the full structured assessment when available. On WSL2, keep `duration` short because PowerShell recording may wait for the full requested duration before analysis begins. ### MCP Apps voice panel Clients that support MCP Apps can call `open_voice_panel` to render the `ui://pronunciation/voice-panel` resource. The panel requests browser microphone access, records locally in the browser, uploads a WAV clip through `analyze_uploaded_audio`, and displays the returned transcript and feedback. The uploaded clip is stored in the same voice session registry as MCP-only recordings, so assistants can call `latest_voice_capture` after the panel finishes and respond to both the development note and the pronunciation feedback. Clients without MCP Apps support should use the visible voice-capture workflow above. ## Prompt Shortcuts MCP clients that expose server prompts can start common workflows directly: | Prompt | Purpose | |---|---| | `start_voice_chat` | Start a local voice conversation with light feedback. | | `daily_practice` | Run a short suggested-sentence practice loop. | | `practice_focus` | Start a drill for a chosen focus and difficulty. | | `troubleshoot_mic` | Inspect microphone devices and recording settings. | ## Configuration ### Whisper model Set `MCP_PRONUNCIATION_MODEL` to pick a different model size: ```bash # Default — fast, English-only (~150 MB) export MCP_PRONUNCIATION_MODEL=base.en # Smaller / faster (~75 MB) export MCP_PRONUNCIATION_MODEL=tiny.en # More accurate (~470 MB) export MCP_PRONUNCIATION_MODEL=small.en # Multilingual options (larger) export MCP_PRONUNCIATION_MODEL=small export MCP_PRONUNCIATION_MODEL=medium ``` Available: `tiny`, `tiny.en`, `base`, `base.en`, `small`, `small.en`, `medium`, `medium.en`, `large-v3`, `large-v3-turbo`. For English-only use, the `.en` variants are faster and more accurate at a given size. GPU (CUDA 12 + cuDNN 9) is auto-detected when available; otherwise runs on CPU with int8 quantization. ### Cache location By default Whisper weights are cached in `~/.cache/huggingface/hub/`. Override with `HF_HUB_CACHE`: ```bash export HF_HUB_CACHE=/path/to/cache ``` ### Startup preload By default the server preloads the Whisper model in the background after the MCP handshake starts. Set `MCP_PRONUNCIATION_PRELOAD=0` for registry inspection, Docker smoke tests, or other environments that only need tool discovery and should avoid model downloads: ```bash export MCP_PRONUNCIATION_PRELOAD=0 ``` ### Temporary recordings Recordings are written as temporary WAV files so `assess` can inspect the last recording. By default they are removed when the server process exits: ```bash export MCP_PRONUNCIATION_AUDIO_RETENTION=session ``` Set `MCP_PRONUNCIATION_AUDIO_RETENTION=keep` if you want temporary recordings to remain on disk for manual inspection. ### Microphone and auto-stop controls By default the server uses your system default microphone. Native `sounddevice` recording stops after 1.5 seconds of detected silence. WSL2 records through Windows PowerShell and may wait for the full requested duration, so use a short `duration` value for quick voice checks. You can override native recording behavior: ```bash # Use a specific input device index or name from the `check_mic` tool export MCP_PRONUNCIATION_INPUT_DEVICE=1 # Options: low, normal, high # high helps soft speakers; low is better in noisy rooms export MCP_PRONUNCIATION_VAD_SENSITIVITY=high # Seconds of silence before auto-stop, clamped to 0.3-5.0 export MCP_PRONUNCIATION_SILENCE_DURATION=2.0 ``` Run `check_mic` to see the default input device, available device indexes, and the active VAD settings. ### Model override in MCP clients ```bash # Codex CLI codex mcp add --env MCP_PRONUNCIATION_MODEL=small.en pronunciation -- uvx mcp-server-pronunciation # Claude Code claude mcp add pronunciation -e MCP_PRONUNCIATION_MODEL=small.en -- uvx mcp-server-pronunciation ``` ### Phoneme analysis extras Installing `mcp-server-pronunciation[phoneme]` enables wav2vec2-based CTC forced alignment. It verifies which reference words the user acoustically produced, regardless of how Whisper's language-model-weighted decoder rewrote them — so rare proper nouns and domain terms no longer surface as false mispronunciations. On first run the extra downloads ~360 MB of weights into `~/.cache/torch/hub/` (override via `TORCH_HOME`). Inference is CPU-only by default and runtime-quantized to int8 (~95 MB RAM). Without the extra, `assess` / `practice` still run the full pipeline except for the forced-alignment step: you get Needleman-Wunsch word alignment against the Whisper hypothesis, CMUdict phoneme-sequence diff, learner-profile hints, and prosody. ## Platform Support | Platform | Recording method | Status | |----------|------------------|--------| | macOS | sounddevice (bundled PortAudio) | Supported | | Linux | sounddevice (needs `libportaudio2`) | Supported | | Windows | sounddevice (bundled PortAudio) | Supported | | WSL2 | PowerShell MCI (winmm.dll) | Supported | **WSL2 note**: WSLg's PulseAudio does not forward microphone audio from the Windows host. This server detects WSL2 automatically and records through PowerShell on the Windows side instead. WSL2 recording may wait for the full requested duration instead of auto-stopping on silence. ## Troubleshooting ### `uvx mcp-server-pronunciation doctor` is your first stop It reports on PortAudio, input devices, Whisper model cache, pronunciation resources, optional forced-alignment dependencies, free disk space, and Python version. Run it whenever something feels off. ### `sounddevice` import fails on Linux You're missing `libportaudio2`. See the install section above. After installing: ```bash uvx mcp-server-pronunciation doctor ``` ### No audio captured / empty recording - **macOS**: System Settings → Privacy & Security → Microphone. Grant access to the app that launched your MCP client, such as Codex CLI, Claude Desktop, or Claude Code. - **Linux**: Check `pavucontrol` (PulseAudio) or `pw-cli list-objects` (PipeWire) for input levels. On PipeWire-only systems, install `pipewire-alsa`. - **WSL2**: Test your mic in Windows Settings → Sound → Input. The server records through Windows, not through WSLg. ### First run is slow The Whisper model downloads on first use (~150 MB for `base.en`). Pre-download it once: ```bash uvx mcp-server-pronunciation pull-model base.en ``` Subsequent runs reuse the cached weights. If startup still feels slow, try `MCP_PRONUNCIATION_MODEL=tiny.en`. ### Claude Desktop on macOS: `spawn uvx ENOENT` Claude Desktop launches MCP servers from a GUI-only environment without `~/.local/bin` on PATH. Use the absolute path to `uvx` in your config (`/Users/YOU/.local/bin/uvx` or wherever `which uvx` reports). ## Known Limitations - This is a stable package release, but the pronunciation and prosody feedback remain experimental coaching signals. Bugs, runtime errors, inaccurate feedback, and platform-specific recording issues can still occur. - Pronunciation scores are coaching signals, not standardized-test, clinical, or native-speaker-equivalence judgments. - Whisper can still mishear rare names, domain terms, short clips, quiet audio, or heavily accented speech. The optional `[phoneme]` extra reduces some reference-sentence false positives but does not eliminate them. - Prosody feedback is heuristic. Pitch tracking can be unreliable with noisy audio, very short utterances, vocal fry, overlapping speech, or clipped recordings. - Learner-profile hints are intentionally rule-based. The current package includes Korean-L1 hints, but they can miss errors, over-trigger on ASR mistakes, and should be treated as targeted practice aids. Contributions for additional L1 profiles are welcome. - First-time setup may download model or pronunciation resources. Run `doctor` and `pull-model` before relying on the server in a live session. - Temporary WAV recordings are written under the system temp directory so that the last recording can be assessed. By default they are removed when the server exits. Set `MCP_PRONUNCIATION_AUDIO_RETENTION=keep` if you want to inspect them later. ## Benchmark Status This project is moving toward benchmark-backed scoring. Planned public benchmark work is tracked in [ROADMAP.md](ROADMAP.md), the testing methodology lives in [docs/TESTING.md](docs/TESTING.md), and the current benchmark helper docs live in [docs/BENCHMARKS.md](docs/BENCHMARKS.md). The primary candidate is Speechocean762 because it has a permissive CC BY 4.0 license and multi-level expert pronunciation scores. L2-ARCTIC is useful for phone-error and learner-profile research checks, including Korean-L1 subset review, but its non-commercial license means it should remain optional and separate from default release claims. ## Publication Status The source repository is public. PyPI, GitHub Release, and MCP Registry publication steps are tracked in [docs/PUBLICATION.md](docs/PUBLICATION.md). ## Privacy - All audio processing happens **locally** on your machine. - Recordings are temporary `.wav` files under your system temp directory (`$TMPDIR`) and are removed when the server exits unless `MCP_PRONUNCIATION_AUDIO_RETENTION=keep` is set. - The Whisper model runs locally — no audio data is sent to any external service. - When the optional `[phoneme]` extra is installed, the wav2vec2 forced aligner also runs locally. Weights are downloaded once from the PyTorch Hub. - No telemetry. No analytics. No network calls except the one-time model weight downloads (Whisper from Hugging Face, wav2vec2 from PyTorch Hub). ## Development ```bash git clone https://github.com/JuhongPark/mcp-server-pronunciation.git cd mcp-server-pronunciation uv sync --extra dev uv run pytest -v uv run ruff check . uv run ruff format --check . ``` To work on the optional wav2vec2 forced-alignment path, install the phoneme extra as well: ```bash uv sync --extra dev --extra phoneme ``` ## Support Issues: https://github.com/JuhongPark/mcp-server-pronunciation/issues ## License MIT. See [LICENSE](LICENSE). Third-party components (all MIT / permissive): - [`faster-whisper`](https://github.com/SYSTRAN/faster-whisper) — MIT - [OpenAI Whisper models](https://github.com/openai/whisper) — MIT - [CTranslate2](https://github.com/OpenNMT/CTranslate2) — MIT - [`sounddevice`](https://python-sounddevice.readthedocs.io/) — MIT - [PortAudio](http://www.portaudio.com/) — MIT - [`cmudict`](https://pypi.org/project/cmudict/) — BSD - [`g2p-en`](https://github.com/Kyubyong/g2p) — Apache 2.0 - [`librosa`](https://librosa.org/) — ISC - Optional (`[phoneme]` extra): [PyTorch](https://pytorch.org/) — BSD, [torchaudio](https://pytorch.org/audio/) — BSD, [wav2vec2 weights](https://github.com/facebookresearch/fairseq) — MIT