chimera-voice 0.1.0

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	✅	✅	⚠️
No model weights	✅	✅	❌
Multi-speaker, per-speaker keys	✅	❌	❌
Automatic VAD (ignores noise)	✅	❌	❌
Natural-sounding output	✅	❌	✅
Deterministic / auditable	✅	❌	❌
Cryptographically one-way	✅	❌	❌
Real-time microphone support	✅	✅	GPU

---

How It Works

Chimera applies six processing stages to every audio input:

Input audio
    │
    ▼  [1] VAD — isolates speech from noise, music, silence
    │
    ▼  [2] Diarization — MFCC k-means, assigns segments to speakers
    │
    ▼  [3] Key Derivation — HKDF-SHA256 per speaker → 8 parameters
    │
    ▼  [4] 7-Layer Vocoder Stack (WORLD)
    │       L1 Pitch shift       F0 × 2^(Δst/12)
    │       L2 Sinusoidal vibrato modulation
    │       L3 Micro-temporal jitter
    │       L4 Formant warp      SP resampled at α·f
    │       L5 Spectral tilt     ±4 dB/kHz ramp
    │       L6 Sub-harmonic injection
    │       L7 Breathiness blend toward noise
    │
    ▼  [5] COWL — Cryptographic One-Way Layer
    │       Sub-perceptual spectral noise (SSNI)
    │       Phase randomisation
    │       Non-linear spectral quantisation (NLSQ)
    │
    ▼  [6] Cross-fade stitching + normalisation
    │
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       : 0.91234×
#   Spectral tilt      : -2.814 dB/kHz
#   Breathiness        : 0.3122
#   Temporal jitter    : 0.01203 σ
#   Vibrato rate       : 4.210 Hz
#   Vibrato depth      : 0.2891 st
#   Subharmonic mix    : 0.0984
#   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. Non-linear μ-law spectral quantisation (ill-posed)
2. Key-seeded phase randomisation (requires HKDF seed)
3. Key-derived sub-perceptual noise injection (requires HMAC sub-key)
4. Seven vocoder transformation layers (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      # 7-layer WORLD vocoder stack
│   ├── 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 (26 tests)
├── 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, all seven vocoder 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 = {0.1.0},
  license = {MIT}
}

---

Changelog

See CHANGELOG.md.

Contributing

See CONTRIBUTING.md.

License

MIT © 2026 Ohswedd