langlearn-tts

A Claude Desktop extension that gives Claude the ability to speak. Ask Claude to pronounce words, generate audio flashcards, or run a full language lesson with audio — in 70+ languages.

Status (2026-02-21)

• Core CLI and MCP surfaces are working with ElevenLabs, OpenAI TTS, and AWS Polly.

• Provider defaults and audio variants (word + example) are still being standardized.

• Claude Desktop directory submission is pending external review.

Roadmap

See ROADMAP.md.

Quick Start

1. Get a TTS API key

You need an account with at least one text-to-speech provider:

• ElevenLabs — best quality, 70+ languages, 5,000+ voices. Free tier: 10K chars/month. (Recommended)

• OpenAI TTS — good quality, easiest setup, 57 languages, 9 voices.

• AWS Polly — better quality, 41 languages, 100+ voices, difficult setup (setup guide).

2. Install in Claude Desktop

Download punt-langlearn-tts.mcpb and double-click to install. Claude Desktop will prompt you for your API key and an output directory.

3. Set up a tutor project (optional)

langlearn-tts ships with 28 tutor prompts — one for each combination of 7 languages and 4 levels. Setting up a project gives Claude a tutor persona that generates audio during lessons.

1. In Claude Desktop, click Projects in the sidebar

2. Click Create Project and name it (e.g., "German with Herr Schmidt")

3. Open the project, click Set custom instructions

4. Copy a prompt from the prompts directory and paste it into the Instructions field

5. Start a new conversation within that project

Each prompt is calibrated to the student's level, based on Mollick & Mollick's "Assigning AI" framework.

4. Try it out

In any Claude Desktop conversation, try:

"Say 'Guten Morgen' in German"

"Create an audio flashcard: 'good morning' in English, then 'Guten Morgen' in German"

"Synthesize these Spanish words as a merged audio file: hola, gracias, por favor, de nada"

"Generate pair flashcards for these German vocabulary words: strong/stark, house/Haus, book/Buch"

Audio plays automatically after each request. Files are saved to your output directory (~/langlearn-audio by default).

Features

• Pronounce anything — ask Claude to say a word or phrase and hear it spoken aloud

• Audio flashcards — Claude creates an MP3 with English first, then the target language, with a pause between them

• Vocabulary lists — give Claude a list of words and get back individual or merged audio files

• 70+ languages — German, Spanish, French, Russian, Korean, Japanese, Chinese, and many more

• Tutor mode — 28 built-in tutor personas that teach with audio throughout the lesson

• Multiple voices — each provider offers a range of voices; ask Claude to use a specific one by name

• Adjustable speed — audio defaults to 90% speed so learners can hear pronunciation clearly

Troubleshooting

If something isn't working, ask Claude to run a health check:

"Run the doctor command to check if everything is set up correctly"

Logs are written to ~/.langlearn-tts/logs/langlearn-tts.log (never contains the text you synthesize). See PRIVACY.md for details.

Developer Reference

Everything below is for developers using the CLI, integrating with other MCP clients, or contributing to the project.

Claude Code / CLI

curl -fsSL https://raw.githubusercontent.com/punt-labs/langlearn-tts/14f4194/install.sh | sh

The default provider is AWS Polly. To use a different provider:

LANGLEARN_TTS_PROVIDER=elevenlabs curl -fsSL https://raw.githubusercontent.com/punt-labs/langlearn-tts/14f4194/install.sh | sh

uv tool install punt-langlearn-tts
langlearn-tts install --provider polly
langlearn-tts doctor

curl -fsSL https://raw.githubusercontent.com/punt-labs/langlearn-tts/14f4194/install.sh -o install.sh
shasum -a 256 install.sh
cat install.sh
sh install.sh

Install ffmpeg for audio stitching (pairs, merged batches):

# macOS (requires Homebrew — install from https://brew.sh if needed)
brew install ffmpeg

# Linux — see https://ffmpeg.org/download.html for your distro

# Windows
winget install --id Gyan.FFmpeg

Claude Desktop setup via CLI

langlearn-tts install

Writes to ~/Library/Application Support/Claude/claude_desktop_config.json (macOS). Options: --provider NAME, --output-dir PATH, --uvx-path PATH. Restart Claude Desktop after running.

Or add manually:

{
  "mcpServers": {
    "langlearn-tts": {
      "command": "/absolute/path/to/uvx",
      "args": ["--from", "punt-langlearn-tts", "langlearn-tts-server"],
      "env": {
        "LANGLEARN_TTS_OUTPUT_DIR": "/absolute/path/to/output/directory"
      }
    }
  }
}

Claude Desktop does not inherit your shell environment. API keys must be literal values (env var references are not supported). Restart after editing.

Environment variables

For Polly, AWS credentials are read from ~/.aws/credentials.

CLI Usage

# Single synthesis
langlearn-tts synthesize "Guten Morgen" --voice daniel -o morning.mp3

# Custom speech rate (percentage, default 90)
langlearn-tts synthesize "Привет" --voice tatyana --rate 70 -o privet.mp3

# ElevenLabs with voice settings
langlearn-tts synthesize "Guten Morgen" --voice Rachel \
  --stability 0.5 --similarity 0.7 --style 0.3 --speaker-boost

# Pair: English + German stitched with a pause
langlearn-tts synthesize-pair "good morning" "Guten Morgen" \
  --voice1 joanna --voice2 daniel -o pair.mp3

# Batch from JSON file (["hello", "world", "good morning"])
langlearn-tts synthesize-batch words.json -d output/

# Batch merged into single file
langlearn-tts synthesize-batch words.json -d output/ --merge --pause 800

# Pair batch from JSON file ([["strong", "stark"], ["house", "Haus"]])
langlearn-tts synthesize-pair-batch pairs.json -d output/

# Browse AI tutor prompts
langlearn-tts prompt list
langlearn-tts prompt show german-high-school | pbcopy

Voices

ElevenLabs — 5,000+ voices. Any voice works with any language. You can also pass a voice ID directly (the 20-character string from the ElevenLabs dashboard). Voice settings: --stability, --similarity, --style (0.0–1.0), --speaker-boost (flag).

AWS Polly — 93 voices from the AWS Polly voice list. Each voice is trained for a specific language. Engine (neural, standard, generative, long-form) is selected automatically.

OpenAI TTS — 9 voices: alloy, ash, coral, echo, fable, onyx, nova, sage, shimmer. Default model: tts-1. Use --model tts-1-hd for higher quality.

All voice names are case-insensitive.

MCP Tools

Each tool accepts auto_play (default: true) to play audio immediately after synthesis.

Other MCP clients

langlearn-tts works with any MCP client that supports stdio transport. Use the server command uvx --from punt-langlearn-tts langlearn-tts-server with the environment variables above. Find your uvx path with which uvx — all paths must be absolute.

Development

git clone https://github.com/punt-labs/langlearn-tts.git
cd langlearn-tts
uv sync --all-extras

uv run pytest tests/ -v
uv run ruff check src/ tests/
uv run ruff format src/ tests/
uv run mypy src/ tests/
uv run pyright src/ tests/

License

MIT