Listenr

Build better speech-to-text and ASR models entirely on your machine.

Record your voice. Clean it up with local AI. Fine-tune Whisper or Moonshine. Deploy something that's actually yours.

Walkthrough  |  Setup  |  Configuration  |  Recording  |  Dataset  |  Troubleshooting

How it works

1. Create good data - Use Listenr to record and collect natural speech with domain-specific vocabulary that generic models miss.
2. Process & improve - Pipe it through Lemonade or any OpenAI-compatible provider to transcribe with Whisper and automatically correct grammar, punctuation, and homophones using a local LLM.
3. Fine-tune & deploy - Use Listenr to build train/dev/test splits and fine-tune Whisper or Moonshine with LoRA. Merge the adapter into a self-contained model you can deploy.

Everything stays local - no audio, text, or weights ever leave on your machine.

Get started

Install Lemonade and pull models:

Lemonade guide: lemonade-server.ai/docs/guide/install

# after installing locally, download default models
lemonade pull Whisper-Base
lemonade pull gpt-oss-20b-mxfp4-GGUF

Install Listenr and start recording:

uv tool install listenr   # or: pipx install listenr
listenr record            # start recording

On PyPI. Python 3.11 or newer. The core install covers recording, transcription and dataset building; fine-tuning and the dataset importers live behind extras, listed in docs/setup.md.

On an AMD GPU, install a ROCm torch before the finetune extra, or pip will resolve a CUDA wheel that cannot see your card. See docs/setup.md.

Working on Listenr itself? Clone the repo and uv pip install -e ".[dev]" instead.

Once you have recordings, process & fine-tune:

# Build train/dev/test splits from your manifest
listenr build-dataset --format hf

# Fine-tune Whisper or Moonshine (see docs/finetune-amd.md for AMD GPUs)
podman compose run --rm finetune

# Merge the LoRA adapter into a standalone model
podman compose run --rm merge

# Evaluate it on the held-out test split
listenr eval --compare-base --keyword YourDomainWord

See docs/setup.md for full installation details.

If you want to mix in an external ASR dataset, use the optional importers to write a separate Listenr-compatible manifest — listenr import-mdc <dataset-id> (Mozilla Data Collective) or listenr import-hf <dataset-id> (Hugging Face) — then pass that manifest to listenr build-dataset alongside your normal one. See docs/dataset.md for details.

Under the hood

Recording & transcription - Listenr streams your microphone to Lemonade's /realtime WebSocket in ~85 ms chunks (16 kHz). Lemonade's voice activity detection segments speech, runs Whisper.cpp, and streams back transcripts.

Auto-correction - A local LLM cleans up punctuation, grammar, and homophones, producing a higher-quality training corpus than raw Whisper output alone.

Dataset & fine-tuning - Listenr saves each utterance as a .wav clip and a line in manifest.jsonl. One command builds train/dev/test splits in HuggingFace format. Another command fine-tunes any openai/whisper-* or UsefulSensors/moonshine-* model using LoRA (works on AMD and NVIDIA GPUs via Podman). Moonshine is the smaller, English-only, edge-oriented option; Whisper is the multilingual all-rounder.

Deployment - listenr merge folds the LoRA adapter into a self-contained model that loads with plain transformers. No PEFT dependency. Run inference locally or deploy it anywhere.

Documentation

Acknowledgments

• Lemonade Server - unified local inference API
• whisper.cpp - fast local ASR
• llama.cpp - fast local LLMs

License

Mozilla Public License Version 2.0 - see LICENSE.