Skip to content

Dependencies (direct) + models/weights

This document lists direct dependencies declared in pyproject.toml, plus the models/weights AbstractVoice may download.

Why direct-only?

  • Direct dependencies (what we declare) are stable and reviewable.
  • Transitive dependencies (what pip resolves) vary by platform and change over time; we call out only a few notable runtimes where it helps debugging.

If you need licensing guidance for model weights and voices, also read docs/voices-and-licenses.md.

Lightweight base dependencies (installed by default)

These are always installed with pip install abstractvoice.

  • numpy
  • Why: array operations and audio sample handling.
  • Where: widely used across abstractvoice/* (audio + cloning glue).
  • Repo: https://github.com/numpy/numpy

  • License: https://github.com/numpy/numpy/blob/main/LICENSE.txt

  • requests

  • Why: HTTP calls (mainly integration / utility flows).
  • Where: various network helpers; also used by some upstream libs.
  • Repo: https://github.com/psf/requests

  • License: https://github.com/psf/requests/blob/main/LICENSE

  • appdirs

  • Why: OS-appropriate cache/data directory locations.
  • Where: used in cache/storage helpers.
  • Repo: https://github.com/ActiveState/appdirs
  • License: https://github.com/ActiveState/appdirs/blob/master/LICENSE.txt

Optional extras (opt-in)

abstractvoice[apple] / abstractvoice[gpu] — platform local stacks

These are the platform local/offline profiles. They include local Piper TTS, Supertonic 3 ONNX TTS, faster-whisper STT, microphone/playback/VAD dependencies, AEC where supported, and the optional local cloning/TTS engine dependencies that are resolver-safe on the current Python version. Use granular extras below for smaller installs, or abstractvoice[all-apple] / abstractvoice[all-gpu] when you also want the web example.

  • piper-tts
  • Why: local neural TTS backend.
  • Where: abstractvoice/adapters/tts_piper.py
  • Repo: https://github.com/rhasspy/piper

  • License: https://github.com/rhasspy/piper/blob/master/LICENSE

  • onnxruntime

  • Why: ONNX inference runtime for Supertonic 3 and Piper execution-provider introspection.
  • Where: abstractvoice/supertonic/runtime.py, abstractvoice/adapters/tts_piper.py
  • Repo: https://github.com/microsoft/onnxruntime

  • License: https://github.com/microsoft/onnxruntime/blob/main/LICENSE

  • faster-whisper

  • Why: local STT backend (fast Whisper inference).
  • Where: abstractvoice/adapters/stt_faster_whisper.py
  • Repo: https://github.com/SYSTRAN/faster-whisper

  • License: https://github.com/SYSTRAN/faster-whisper/blob/master/LICENSE

  • sounddevice

  • Why: microphone/speaker I/O (PortAudio bindings).
  • Where: abstractvoice/recognition.py, abstractvoice/tts/tts_engine.py, abstractvoice/audio/*
  • Repo: https://github.com/spatialaudio/python-sounddevice

  • License: https://github.com/spatialaudio/python-sounddevice/blob/master/LICENSE

  • soundfile

  • Why: WAV/FLAC/OGG read/write (libsndfile bindings).
  • Where: cloning store normalization, tests, file utilities.
  • Repo: https://github.com/bastibe/python-soundfile

  • License: https://github.com/bastibe/python-soundfile/blob/master/LICENSE

  • webrtcvad

  • Why: VAD (voice activity detection) for listen() and voice modes.
  • Where: abstractvoice/vad/voice_detector.py, abstractvoice/recognition.py
  • Repo: https://github.com/wiseman/py-webrtcvad
  • License: https://github.com/wiseman/py-webrtcvad/blob/master/LICENSE.txt

abstractvoice[supertonic] — Supertonic 3 fixed-profile ONNX TTS

Supertonic is local TTS only, not voice cloning. AbstractVoice integrates the published ONNX files through an internal adapter and does not depend on the external supertonic Python SDK.

Python packages:

  • onnxruntime
  • Why: executes the Supertonic ONNX graphs.
  • Where: abstractvoice/supertonic/runtime.py, abstractvoice/adapters/tts_supertonic.py
  • Repo: https://github.com/microsoft/onnxruntime
  • License: https://github.com/microsoft/onnxruntime/blob/main/LICENSE

Models/weights:

  • Supertone Supertonic 3
  • Model: https://huggingface.co/Supertone/supertonic-3
  • License file: https://huggingface.co/Supertone/supertonic-3/blob/main/LICENSE
  • Cache: ~/.cache/abstractvoice/supertonic-3 by default
  • Prefetch: python -m abstractvoice download --supertonic or abstractvoice-prefetch --supertonic
  • Profiles: built-in fixed styles M1-M5 and F1-F5

abstractvoice[stt-hf] — local Transformers ASR (Hugging Face)

This extra enables the stt_engine="transformers-asr" path for local speech-to-text using Hugging Face models (for example Whisper and Qwen3-ASR).

Python packages:

  • torch
  • Why: tensor runtime for local ASR inference.
  • Where: abstractvoice/adapters/stt_transformers_asr.py
  • Repo: https://github.com/pytorch/pytorch

  • License: https://github.com/pytorch/pytorch/blob/main/LICENSE

  • transformers

  • Why: HF model + processor APIs.
  • Where: abstractvoice/adapters/stt_transformers_asr.py
  • Repo: https://github.com/huggingface/transformers

  • License: https://github.com/huggingface/transformers/blob/main/LICENSE

  • soundfile

  • Why: audio file I/O (transcribe() and byte decoding).
  • Where: abstractvoice/adapters/stt_transformers_asr.py
  • Repo: https://github.com/bastibe/python-soundfile
  • License: https://github.com/bastibe/python-soundfile/blob/master/LICENSE

Models/weights (Hugging Face):

  • Whisper (Transformers)
  • Example ids: openai/whisper-large-v3, openai/whisper-large-v3-turbo
  • Cache: ~/.cache/huggingface (default)

  • Prefetch: python -m abstractvoice download --stt-hf openai/whisper-large-v3 or abstractvoice-prefetch --stt-hf openai/whisper-large-v3

  • Qwen3-ASR-1.7B

  • Model: https://huggingface.co/Qwen/Qwen3-ASR-1.7B
  • Cache: ~/.cache/huggingface (default)
  • Prefetch: python -m abstractvoice download --stt-hf Qwen/Qwen3-ASR-1.7B or abstractvoice-prefetch --stt-hf Qwen/Qwen3-ASR-1.7B

Vendored code:

  • Qwen3-ASR Transformers backend
  • Upstream: qwen-asr (Apache-2.0) and https://github.com/QwenLM/Qwen3-ASR
  • What we ship: a derived Transformers implementation under abstractvoice/qwen3_asr/* so Qwen3-ASR works without trust_remote_code.

abstractvoice[omnivoice], [cloning], [chroma], [audiodit] — model prefetch support

  • huggingface_hub
  • Why: download and cache optional engine model weights/artifacts.
  • Where: abstractvoice/cloning/engine_f5.py, abstractvoice/cloning/engine_chroma.py, abstractvoice/audiodit/runtime.py, abstractvoice/omnivoice/runtime.py
  • Repo: https://github.com/huggingface/huggingface_hub
  • License: https://github.com/huggingface/huggingface_hub/blob/main/LICENSE

abstractvoice[audiodit] — LongCat-AudioDiT (TTS + prompt-audio cloning)

Supported on Python 3.9+. Python 3.9 uses the latest compatible Transformers 4.x line; Python 3.10+ uses the Transformers 5.x line.

Python packages:

  • torch
  • Why: tensor runtime for AudioDiT inference.
  • Where: abstractvoice/audiodit/*, abstractvoice/cloning/engine_audiodit.py
  • Repo: https://github.com/pytorch/pytorch

  • License: https://github.com/pytorch/pytorch/blob/main/LICENSE

  • transformers

  • Why: HF PreTrainedModel/tokenizer APIs used by AudioDiT + text encoder.
  • Where: abstractvoice/audiodit/*, abstractvoice/audiodit/runtime.py
  • Repo: https://github.com/huggingface/transformers

  • License: https://github.com/huggingface/transformers/blob/main/LICENSE

  • einops

  • Why: model block reshaping utilities (used by AudioDiT modules).
  • Where: abstractvoice/audiodit/modeling_audiodit.py
  • Repo: https://github.com/arogozhnikov/einops

  • License: https://github.com/arogozhnikov/einops/blob/master/LICENSE

  • sentencepiece

  • Why: tokenizer runtime (used by UMT5 tokenizer).
  • Where: pulled by Transformers tokenizers depending on model.
  • Repo: https://github.com/google/sentencepiece

  • License: https://github.com/google/sentencepiece/blob/master/LICENSE

  • safetensors

  • Why: weight format used by HF models.
  • Where: model load via Transformers.
  • Repo: https://github.com/huggingface/safetensors
  • License: https://github.com/huggingface/safetensors/blob/main/LICENSE

Models/weights (Hugging Face):

  • LongCat-AudioDiT-1B (weights + model license statement)
  • Model: https://huggingface.co/meituan-longcat/LongCat-AudioDiT-1B
  • License file: https://huggingface.co/meituan-longcat/LongCat-AudioDiT-1B/blob/main/LICENSE
  • Cache: ~/.cache/huggingface (default)

  • Prefetch: python -m abstractvoice download --audiodit or abstractvoice-prefetch --audiodit

  • UMT5 text encoder (default: google/umt5-base)

  • Model: https://huggingface.co/google/umt5-base
  • License: https://huggingface.co/google/umt5-base (model card metadata; license is apache-2.0)
  • Cache: ~/.cache/huggingface (default)

Vendored code:

  • LongCat-AudioDiT upstream repository (MIT)
  • Repo: https://github.com/meituan-longcat/LongCat-AudioDiT
  • License: https://github.com/meituan-longcat/LongCat-AudioDiT/blob/main/LICENSE
  • What we ship: a HuggingFace-compatible derived implementation under abstractvoice/audiodit/* to avoid trust_remote_code.

abstractvoice[omnivoice] — OmniVoice (default local cloning + omnilingual TTS / design)

Requires Python 3.10+ because the upstream omnivoice package declares that Python floor.

OmniVoice is the recommended/default local cloning backend for new AbstractVoice clones. Install this extra, or a platform/full profile that includes it, before using clone commands without --engine.

Python packages:

  • omnivoice
  • Why: OmniVoice engine runtime + model wrapper.
  • Where: abstractvoice/omnivoice/*, abstractvoice/adapters/tts_omnivoice.py, abstractvoice/cloning/engine_omnivoice.py
  • Repo: https://github.com/k2-fsa/OmniVoice

  • License: https://github.com/k2-fsa/OmniVoice/blob/master/LICENSE

  • torch / torchaudio / torchvision

  • Why: tensor runtime for OmniVoice inference.
  • Repo: https://github.com/pytorch/pytorch

  • License: https://github.com/pytorch/pytorch/blob/main/LICENSE

  • transformers

  • Repo: https://github.com/huggingface/transformers

  • License: https://github.com/huggingface/transformers/blob/main/LICENSE

  • accelerate

  • Repo: https://github.com/huggingface/accelerate
  • License: https://github.com/huggingface/accelerate/blob/main/LICENSE

Models/weights (Hugging Face):

  • OmniVoice (weights + model license statement)
  • Model: https://huggingface.co/k2-fsa/OmniVoice
  • Cache: ~/.cache/huggingface (default)
  • Prefetch: python -m abstractvoice download --omnivoice or abstractvoice-prefetch --omnivoice

abstractvoice[chroma] — Chroma-4B cloning (GPU-heavy)

Requires Python 3.10+ in AbstractVoice packaging because Chroma depends on Transformers 5.x.

  • torch / torchaudio / torchvision
  • Repo: https://github.com/pytorch/pytorch

  • License: https://github.com/pytorch/pytorch/blob/main/LICENSE

  • transformers

  • Repo: https://github.com/huggingface/transformers

  • License: https://github.com/huggingface/transformers/blob/main/LICENSE

  • accelerate

  • Repo: https://github.com/huggingface/accelerate

  • License: https://github.com/huggingface/accelerate/blob/main/LICENSE

  • av (PyAV)

  • Repo: https://github.com/PyAV-Org/PyAV

  • License: https://github.com/PyAV-Org/PyAV/blob/main/LICENSE.txt

  • librosa / audioread / pillow / safetensors

  • librosa repo/license: https://github.com/librosa/librosa / https://github.com/librosa/librosa/blob/main/LICENSE.md
  • audioread repo/license: https://github.com/beetbox/audioread / https://github.com/beetbox/audioread/blob/master/LICENSE
  • pillow repo/license: https://github.com/python-pillow/Pillow / https://github.com/python-pillow/Pillow/blob/main/LICENSE
  • safetensors repo/license: https://github.com/huggingface/safetensors / https://github.com/huggingface/safetensors/blob/main/LICENSE

Model:

  • Chroma-4B
  • Model: https://huggingface.co/FlashLabs/Chroma-4B
  • License: see model card + repo files.

abstractvoice[cloning] — OpenF5 / F5-TTS cloning

Requires Python 3.10+ because current upstream f5-tts packages import Python 3.10-only annotations during inference. On Python 3.9, use abstractvoice[audiodit] with the audiodit cloning engine for tested prompt-audio cloning.

This extra remains the explicit OpenF5 backend. It is not the default local cloning path; new clones default to OmniVoice unless you pass engine="f5_tts" or --engine f5_tts.

  • f5-tts
  • Why: cloning engine (f5_tts) backend.
  • Where: abstractvoice/cloning/engine_f5.py
  • Repo: https://github.com/SWivid/F5-TTS
  • License: https://github.com/SWivid/F5-TTS/blob/main/LICENSE

abstractvoice[audio-fx] — Audio effects

  • librosa
  • Why: optional “speed change without pitch change”.
  • Where: used by the audio FX path; surfaced via docs/installation.md.
  • Repo: https://github.com/librosa/librosa
  • License: https://github.com/librosa/librosa/blob/main/LICENSE.md

abstractvoice[aec] — Acoustic echo cancellation

Requires Python 3.11+ because aec-audio-processing declares that Python floor.

  • aec-audio-processing
  • Why: optional AEC for “full voice” barge-in on speakers.
  • Where: abstractvoice/aec/*
  • Repo: https://github.com/shichaog/AEC-Audio-Processing
  • License: https://github.com/shichaog/AEC-Audio-Processing/blob/main/LICENSE

abstractvoice[web] — Local web example

The base web extra installs only the browser server stack. Compose it with abstractvoice[all-apple] / abstractvoice[all-gpu] for the full platform local lab, with abstractvoice[supertonic] for the browser UI plus Supertonic, or with other granular engine extras such as abstractvoice[omnivoice] for smaller installs.

  • FastAPI
  • Why: local browser example routes for TTS, transcription, assistant/user voice selection, voice cloning upload forms, and discussion playback (abstractvoice web).
  • Where: abstractvoice/examples/web_ui.py
  • Repo: https://github.com/fastapi/fastapi

  • License: https://github.com/fastapi/fastapi/blob/master/LICENSE

  • Uvicorn

  • Why: local ASGI server for abstractvoice web.
  • Where: abstractvoice/examples/web_ui.py
  • Repo: https://github.com/encode/uvicorn

  • License: https://github.com/encode/uvicorn/blob/master/LICENSE.md

  • python-multipart

  • Why: parse browser/OpenAI-compatible audio upload forms for transcription and voice cloning.
  • Where: abstractvoice/examples/web_ui.py
  • Repo: https://github.com/Kludex/python-multipart
  • License: https://github.com/Kludex/python-multipart/blob/master/LICENSE.txt

abstractvoice[stt] — Faster-Whisper STT

  • faster-whisper
  • Why: local STT adapter path.
  • Where: abstractvoice/adapters/stt_faster_whisper.py
  • Repo: https://github.com/SYSTRAN/faster-whisper
  • License: https://github.com/SYSTRAN/faster-whisper/blob/master/LICENSE

Notable runtimes (debugging-oriented)

  • CTranslate2 (used by faster-whisper)
  • Repo: https://github.com/OpenNMT/CTranslate2

  • License: https://github.com/OpenNMT/CTranslate2/blob/master/LICENSE

  • ONNX Runtime (used directly by Supertonic and inspected by Piper)

  • Repo: https://github.com/microsoft/onnxruntime
  • License: https://github.com/microsoft/onnxruntime/blob/main/LICENSE

Development dependencies

These are used for local development only (not required at runtime).

  • pytest: https://github.com/pytest-dev/pytest — https://github.com/pytest-dev/pytest/blob/main/LICENSE
  • black: https://github.com/psf/black — https://github.com/psf/black/blob/main/LICENSE
  • flake8: https://github.com/PyCQA/flake8 — https://github.com/PyCQA/flake8/blob/main/LICENSE