chimera-voice 0.2.0

Cryptographically irreversible, speaker-aware voice anonymisation

Chimera

Cryptographically Irreversible, Speaker-Aware Voice Anonymisation

Chimera disguises the identity of one or more speakers in any audio recording. It runs entirely on CPU, requires no cloud services, no GPU, and no pre-trained model weights. Given only the output audio and any key, no known algorithm can recover the original speaker's acoustic identity.

Wiki · Quick Start · Paper

---

Why Chimera?

Feature	Chimera	Pitch shift	Neural VC
Fully local / offline	:white_check_mark:	:white_check_mark:	:warning:
No model weights	:white_check_mark:	:white_check_mark:	:x:
Multi-speaker, per-speaker keys	:white_check_mark:	:x:	:x:
Automatic VAD (ignores noise)	:white_check_mark:	:x:	:x:
Natural-sounding output	:white_check_mark:	:x:	:white_check_mark:
Deterministic / auditable	:white_check_mark:	:x:	:x:
Cryptographically one-way	:white_check_mark:	:x:	:x:
Real-time microphone support	:white_check_mark:	:white_check_mark:	GPU

---

How It Works

Chimera applies six processing stages to every audio input:

Input audio
    |
    v  [1] VAD -- isolates speech from noise, music, silence
    |
    v  [2] Diarization -- MFCC k-means, assigns segments to speakers
    |
    v  [3] Key Derivation -- HKDF-SHA256 per speaker -> 5 parameters
    |
    v  [4] Voice Transform (Praat LPC source-filter model)
    |       Pitch shift     via PSOLA resynthesis
    |       Formant shift   via LPC vocal-tract length scaling
    |
    v  [5] COWL -- Cryptographic One-Way Layer
    |       Sub-perceptual spectral noise (SSNI)
    |       Micro phase perturbation (MPP)
    |
    v  [6] Cross-fade stitching + level matching
    |
Output audio (mono float64, same sample rate)

All randomness is derived from the key via HKDF, making the transformation fully deterministic and fully one-way: same key -> same output; output -> original is computationally infeasible.

---

Installation

pip install chimera-voice

With real-time microphone support:

pip install "chimera-voice[realtime]"

From source:

git clone https://github.com/Ohswedd/chimera
cd chimera
pip install -e ".[dev]"

---

Quick Start

Anonymise a file

import chimera

chimera.mask_file("interview.wav", "anonymous.wav", key="my-secret", preset="strong")

Work with NumPy arrays

import chimera
import soundfile as sf

audio, sr = sf.read("interview.wav")
result = chimera.mask_array(audio, sr, key="my-secret", preset="strong")
sf.write("anonymous.wav", result.audio, sr)

Multi-speaker: independent key per speaker

result = chimera.mask_array(
    audio, sr,
    key        = "my-secret",
    preset     = "moderate",
    mode       = chimera.MaskMode.ALL_UNIQUE,   # default
    n_speakers = 3,
)
print(result.speakers_masked)    # ['SPEAKER_0', 'SPEAKER_1', 'SPEAKER_2']
print(f"Processed in {result.processing_time_s:.2f}s")

Mask only selected speakers

result = chimera.mask_array(
    audio, sr,
    key          = "my-secret",
    preset       = "strong",
    mode         = chimera.MaskMode.SELECTED,
    speaker_ids  = ["SPEAKER_0"],   # only mask speaker 0
)

Real-time microphone

from chimera.realtime import RealtimeAnonymiser

anon = RealtimeAnonymiser(key="my-secret", preset="moderate")
anon.start()
input("Recording -- press Enter to stop...")
anon.stop()
anon.save("recorded_anonymous.wav")

Streaming (generator-based)

from chimera.realtime import mask_stream

for masked_chunk in mask_stream(my_chunk_generator, sr=22050,
                                key="my-secret", preset="strong"):
    send_to_output(masked_chunk)

Inspect parameters

p = chimera.get_params("my-secret", preset="strong")
print(p.summary())

#        Chimera MaskParams
# --------------------------------------------
#   Speaker label      : (none)
#   Pitch shift        : +5.743 st
#   Formant warp       : 0.91234x
#   Spectral tilt      : -1.214 dB/kHz
#   Breathiness        : 0.0812
#   Master intensity   : 0.780
# --------------------------------------------

---

Presets

Preset	Intensity	ASV EER	Use case
whisper	0.12	~5 %	Soft watermarking
subtle	0.28	~15 %	Light disguise
moderate	0.52	~35 %	Speaker unrecognisable to humans
strong	0.78	~48 %	ASV systems fail
extreme	1.00	~50 %	Maximum -- content only

Indicative Equal Error Rate against ECAPA-TDNN (VoicePrivacy 2024 protocol).

---

MaskMode

Mode	Behaviour
MaskMode.ALL_UNIQUE	Each speaker gets independent parameters (default)
MaskMode.ALL_SAME	All speakers share the same transformation
MaskMode.SELECTED	Only speakers listed in speaker_ids are masked

---

Security

Chimera is designed with three security properties:

Determinism -- F(audio, key) always returns the same output. Every bit of randomness is derived from the key via HKDF-SHA256.

One-way -- Given F(audio, key), recovering the original speaker's acoustic identity requires simultaneously inverting:

1. Key-seeded micro phase perturbation (requires HKDF seed)
2. Key-derived sub-perceptual noise injection (requires HMAC sub-key)
3. Praat LPC pitch + formant transform (non-invertible without all params)

Key independence -- HKDF-SHA256 provides 128-bit second-preimage resistance. Per-speaker keys are domain-separated: key + ":chimera:spk:" + speaker_id.

Chimera is a privacy-enhancing tool, not an encryption scheme. For high-stakes deployments, combine it with access control, key rotation, and additional anonymisation measures. See the Security wiki page for the full threat model.

---

Supported Audio Formats

Any format supported by libsndfile: WAV, FLAC, OGG, AIFF, and more.

Supported sample rates: 8 000, 16 000, 22 050, 24 000, 44 100, 48 000 Hz.

---

Project Structure

chimera/
|-- chimera/              # Library source
|   |-- __init__.py       # Public API surface
|   |-- core.py           # High-level functions: mask_file, mask_array, get_params
|   |-- pipeline.py       # ChimeraPipeline -- full orchestration
|   |-- keygen.py         # HKDF-SHA256 parameter derivation
|   |-- vad.py            # Voice Activity Detector
|   |-- diarize.py        # MFCC k-means speaker diarizer
|   |-- transform.py      # Praat LPC source-filter voice transform
|   |-- irreversible.py   # Cryptographic One-Way Layer (COWL)
|   |-- realtime.py       # Real-time microphone / streaming engine
|   |-- presets.py        # Named intensity presets
|   |-- types.py          # MaskParams, ChimeraResult, MaskMode, ...
|   |-- exceptions.py     # Exception hierarchy
|   `-- py.typed          # PEP 561 marker
|-- tests/                # Full test suite
|-- examples/             # Runnable usage examples
|-- docs/                 # Documentation
|-- benchmarks/           # Performance benchmarks
|-- paper/                # Academic paper (PDF)
|-- .github/workflows/    # CI and release workflows
|-- pyproject.toml
|-- CHANGELOG.md
|-- CONTRIBUTING.md
`-- LICENSE

---

Development

Setup

git clone https://github.com/Ohswedd/chimera
cd chimera
python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

Run tests

pytest                         # all tests
pytest -m "not slow"           # skip slow tests
pytest --cov=chimera           # with coverage

Lint and format

black chimera tests examples
ruff check chimera tests examples
mypy chimera

Build distribution

pip install build
python -m build

---

Academic Paper

The full technical paper is available at paper/chimera_paper.pdf.

It covers the full threat model (A1-A4), HKDF key derivation, voice transformation layers, COWL security argument, diarization architecture, real-time latency profile, comparison with state-of-the-art, and 15 references.

Citation:

@software{chimera2026,
  title   = {Chimera: Cryptographically Irreversible Speaker-Aware Voice Anonymisation},
  author  = {Ohswedd},
  year    = {2026},
  url     = {https://github.com/Ohswedd/chimera},
  version = {0.2.0},
  license = {MIT}
}

---

Changelog

See CHANGELOG.md.

Contributing

See CONTRIBUTING.md.

License

MIT 2026 Ohswedd