Speechloom

Local audio and video transcription using NeMo-Speech.cpp and Parakeet TDT 0.6B v3. Outputs JSON, plain text, SRT, and WebVTT.

Requirements

• Python 3.10+
• FFmpeg and FFprobe
• A supported native runtime profile for managed setup
• Git, a C++17 compiler, Ninja, and CMake 3.26–3.x only for source builds
• CUDA and nvcc only for CUDA source builds

Setup

Install the Python package from PyPI:

pipx install speechloom
# or: python3 -m pip install speechloom

Install the runtime and ASR model:

speechloom setup

Speechloom selects a usable backend from the capabilities available on the host. You can also select one explicitly:

speechloom setup --backend cuda

To install translation support (about 16 GiB of free space is needed during conversion):

speechloom setup --backend cuda --features translation

Speaker diarization uses the pinned four-speaker Sortformer model:

speechloom setup --features diarization
speechloom transcribe meeting.mp4 --diarize

Multiple optional features can be selected with --features translation,diarization.

Setup uses the platform's standard per-user configuration, data, and cache directories. Existing repository-local .runtime assets are imported in place; they are not moved or deleted.

Portable installations can set SPEECHLOOM_CONFIG_HOME, SPEECHLOOM_DATA_HOME, and SPEECHLOOM_CACHE_HOME explicitly.

Check setup state or remove setup caches with:

speechloom setup status
speechloom setup clean --all

The repository-local setup remains available for development and troubleshooting.

Usage

Check the installation:

speechloom doctor

Transcribe one or more files:

speechloom transcribe recording.mp4
speechloom transcribe recordings/ --recursive --workers 2
speechloom transcribe recording.mp4 --output-dir ./output

To translate a transcript, install the translation feature and provide the source and target languages:

speechloom transcribe russian.mp4 --source-language ru --translate-to en

The original files remain transcript.* and subtitles.*. Translated files are named translation.en.* and subtitles.en.*.

Inspect a completed job:

speechloom inspect transcripts/<job-directory>

Each job contains a manifest, canonical transcript.json, and the requested text and subtitle formats. Completed jobs are reused unless --force is set.

Run speechloom transcribe --help for all options.

Local API

Install the optional server dependencies and allow the directories a desktop client may submit:

python3 -m pip install -e ".[api]"
speechloom serve --allow-root /path/to/media

The API listens on 127.0.0.1:8765; OpenAPI documentation is available at /docs. Remote binding requires --allow-remote and a SPEECHLOOM_API_TOKEN bearer token.

Configuration

Settings are read in this order:

1. command-line options
2. SPEECHLOOM_* environment variables
3. the selected INI file
4. defaults

The default file is in the platform's standard user configuration directory. --config still selects an explicit file. See config.example.ini for available settings.

Tests

PYTHONDONTWRITEBYTECODE=1 PYTHONPATH=src python3 -m unittest discover -s tests -v

Real-runtime tests are enabled by setting SPEECHLOOM_TEST_NEMO, SPEECHLOOM_TEST_MODEL, and SPEECHLOOM_TEST_MEDIA.

License

The project is licensed under MIT. Models are distributed separately under their respective licenses.