Voices in Your Head

Start a thought, then pause. Let the voices continue your thought.

Speech recognition, text generation, and voice synthesis run locally in the browser. Models download once and are cached for future visits.

Run

Requires a WebGPU-capable browser, microphone access, and a recent Node.js version.

npm install
npm run dev

Open localhost:3000/voices. The initial model download is about 1.07 GB; cloning may load additional assets.

Production: npm run build then npm run start.

How it works

Microphone → Silero VAD → Parakeet → LFM2.5 → Pocket TTS → Web Audio

Next.js and React provide the interface; Transformers.js loads recognition and completion models in separate workers. app/llm/LLMPage.tsx coordinates capture, loading, generation, and playback.

• Pauses: Silero detects speech and triggers completion after roughly 480 ms of quiet. Listening is suppressed during playback to avoid hearing the app itself.
• Completions: the model continues your sentence, with a 32-token cap and cleanup of repetition and chatbot-style output.
• Voices: Pocket encodes a recording of up to 15 seconds and reuses its voice conditioning. Synthesis runs sequentially, while chorus tracks can overlap in playback.
• Memory: optional conversation context retains up to eight turns in session memory; Reset clears it.
• Loading: model sessions initialize sequentially with overlapping asset prefetch. Browser caches avoid repeat downloads, and a Web Lock prevents duplicate model stacks in same-origin tabs.

The public model is maxime/personal-notebooks-350m, trained on English translations of Weil's and Camus's notebooks. See its model repository for license terms. Source texts and training data are not published here.

Fine-tuning

Both notebook and personal-notes workflows use LiquidAI/LFM2.5-350M with rank-8 LoRA. Training runs locally on Apple Silicon with MLX-LM and Transformers, outside the browser. Download the base model first:

hf download LiquidAI/LFM2.5-350M config.json generation_config.json tokenizer.json tokenizer_config.json chat_template.jinja model.safetensors --local-dir data/models/lfm2.5-350m-base

For personal Markdown notes:

python scripts/finetune/prepare_dataset.py --notes /path/to/notes
mlx_lm.lora -c scripts/finetune/lora-config.yaml

Preparation uses the runtime prompt and Liquid tokenizer, preserves complete sentence tails, and enforces output/context limits. Use --exclude patterns to omit copied material. Outputs go to ignored data/personal-lfm/. The personal LFM model needs a new training run before use.

For the public notebook workflow, supply the source EPUBs under data/, run scripts/finetune/prepare_notebooks.py, then train with scripts/finetune/notebooks-lora.yaml. Use evaluate_notebooks.py to compare base and tuned generations.

Export personal weights using the Liquid4All onnx-export checkout at data/tools/onnx-export, with its own environment prepared via uv sync --no-dev:

mlx_lm.fuse --model data/models/lfm2.5-350m-base --adapter-path data/personal-lfm/adapters --save-path data/personal-lfm/fused
data/tools/onnx-export/.venv/bin/python scripts/finetune/restore_lfm_hf_layout.py --base data/models/lfm2.5-350m-base --fused data/personal-lfm/fused
data/tools/onnx-export/.venv/bin/lfm2-export data/personal-lfm/fused --output-dir data/personal-lfm --output-name personal-350m --precision q4

Do not skip convolution-layout restoration between MLX fusion and ONNX export. Inspect generated continuations and check for memorization before using an export.

Copy data/personal-lfm/exports/personal-350m/ into public/models/personal-350m/, keeping configs, tokenizer, and ONNX graph/external weights together. Open /voices?debugLlmModel=personal-350m on localhost. Personal weights remain local and unpublished. For notebook exports, use the corresponding data/notebooks/ paths and notebooks-350m name.