mcp-server-pronunciation

Public beta notice

This project is an early beta and is still under active development. It may contain bugs, runtime errors, inaccurate transcripts, inaccurate pronunciation feedback, or platform-specific recording issues. Use it for experimentation and language-learning practice only, and review outputs carefully before relying on them. See DISCLAIMER.md.

An MCP (Model Context Protocol) server that lets you talk to Claude 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 Claude Desktop, Claude Code, and any other MCP client. Everything runs locally — audio is captured with your mic, transcribed by 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, Claude replies, and feedback on what you just said (pronunciation, grammar, fluency) surfaces inside the same tool call so Claude can weave it into a natural reply — or stay out of the way when you're just chatting.

Features

• Voice conversation with Claude. Speak, auto-stop on silence, Claude reads your transcript and responds.

• 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

Public beta

Beta releases are pre-releases. Install the current beta explicitly:

uvx mcp-server-pronunciation@0.3.0b3

For pip users:

pip install --pre mcp-server-pronunciation

Run doctor before relying on the beta in a live session:

mcp-server-pronunciation doctor

General install commands

# 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:

# 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 Claude, run the preflight:

uvx mcp-server-pronunciation doctor

Optional — pre-download the Whisper model (~150 MB) so the first call is instant:

uvx mcp-server-pronunciation pull-model base.en

Add to your MCP client

Claude Code

claude mcp add pronunciation -- uvx mcp-server-pronunciation

Claude Desktop

Edit claude_desktop_config.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:

{
  "mcpServers": {
    "pronunciation": {
      "command": "uvx",
      "args": ["mcp-server-pronunciation"]
    }
  }
}

VS Code (with MCP support)

Add to .vscode/mcp.json or your user settings:

{
  "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."

Claude (calls converse): records your speech, transcribes it, notes that you said "buyed" instead of "bought"

Claude: "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."

Claude (calls suggest_sentence with focus=th): "Try this: The three brothers thought thoroughly about their future."

You: "Record me reading it."

Claude (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."

Claude (calls retry): re-records the same target sentence and compares

Tools

Configuration

Whisper model

Set MCP_PRONUNCIATION_MODEL to pick a different model size:

# 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:

export HF_HUB_CACHE=/path/to/cache

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:

export MCP_PRONUNCIATION_AUDIO_RETENTION=session

Set MCP_PRONUNCIATION_AUDIO_RETENTION=keep if you want temporary recordings to remain on disk for manual inspection.

Model override in 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

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.

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:

uvx mcp-server-pronunciation doctor

No audio captured / empty recording

• macOS: System Settings → Privacy & Security → Microphone. Grant access to the app that launched Claude Desktop / 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:

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 public beta. Bugs, runtime errors, inaccurate feedback, and platform-specific recording issues are expected.

• 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, the testing methodology lives in docs/TESTING.md, and the current benchmark helper docs live in 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.

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

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:

uv sync --extra dev --extra phoneme

Support

Issues: https://github.com/JuhongPark/mcp-server-pronunciation/issues

License

MIT. See LICENSE.

Third-party components (all MIT / permissive):

• faster-whisper — MIT

• OpenAI Whisper models — MIT

• CTranslate2 — MIT

• sounddevice — MIT

• PortAudio — MIT

• cmudict — BSD

• g2p-en — Apache 2.0

• librosa — ISC

• Optional ([phoneme] extra): PyTorch — BSD, torchaudio — BSD, wav2vec2 weights — MIT