Source-linked AI summary
Speaker anonymisation using the McAdams coefficient
Jose Patino, Natalia Tomashenko, Massimiliano Todisco, Andreas Nautsch, Nicholas Evans
TL;DR
Speech anonymisation seeks to reduce the reliability of automatic speaker recognition while preserving intelligibility and naturalness. This paper transforms speech spectral envelopes with the McAdams coefficient and reports a ten-fold increase in baseline equal error rate with modest intelligibility degradation, including against a semi-informed adversary.
Problem
Speech anonymisation must reduce automatic speaker-recognition reliability while preserving intelligibility and naturalness.
Source: §1 · p. 1Method
The approach shifts speech formant positions with the McAdams coefficient and samples random coefficients from ranges to produce different pseudovoices.
Source: §3.1 · p. 2, §3.2 · p. 3Results
The McAdams approach increases the baseline automatic speaker-recognition equal error rate ten-fold, with only modest intelligibility degradation even against a semi-informed adversary.
Source: §5 · p. 4Takeaways & Limitations
The results show that well-known signal-processing techniques can provide efficient and effective anonymisation for speech signals.
Source: §5 · p. 4The work remains far from true anonymisation until error rates reach random performance or ZEBRA category 0, and averaged privacy results can obscure database, gender, and individual differences.
Source: §5 · p. 4
Abstract
from arXiv · showhide
Anonymisation has the goal of manipulating speech signals in order to degrade the reliability of automatic approaches to speaker recognition, while preserving other aspects of speech, such as those relating to intelligibility and naturalness. This paper reports an approach to anonymisation that, unlike other current approaches, requires no training data, is based upon well-known signal processing techniques and is both efficient and effective. The proposed solution uses the McAdams coefficient to transform the spectral envelope of speech signals. Results derived using common VoicePrivacy 2020 databases and protocols show that random, optimised transformations can outperform competing solutions in terms of anonymisation while causing only modest, additional degradations to intelligibility, even in the case of a semi-informed privacy adversary.
1. Introduction
Speech anonymisation aims to reduce speaker-identification reliability while preserving speech utility and voice distinctiveness. This paper explores whether optimised McAdams-based signal processing can improve on a simple baseline and reduce the need for more complex solutions.
1. Introduction
- Speech anonymisation seeks to conceal personally identifiable information in speech, including biometric information, while preserving intelligibility, naturalness, and voice distinctiveness.The VoicePrivacy requirements also call for degrading automatic speaker recognition while retaining utility for applications such as automatic speech recognition.Source: §1 · p. 1
- Evaluation across attack scenarios uses separately configured anonymisation for enrolment and test data so each speaker has different anonymised voices in the two sets.The scenarios vary an adversary's knowledge of or access to a similarly, but not identically, configured system.Source: §1 · p. 1
- The original secondary McAdams baseline is efficient and requires no training data, but was deterministic and unoptimised and preserved voice distinctiveness better than the primary baseline.The primary x-vector and neural-waveform baseline better suppressed speaker-specific information but required substantial training data and computational resources.Source: §1 · p. 1
- This paper optimises the McAdams-based solution and makes it stochastic to improve protection from reversibility while assessing the need for more complex approaches.The authors also hope that combining techniques from the two baselines may offer future opportunities to improve performance.Source: §1 · p. 1
2. Previous work
Previous work connects the McAdams coefficient to additive synthesis and places spectral warping alongside other speech-processing techniques. These approaches differ in whether they modify only the spectral envelope or also pitch.
2. Previous work
- Additive synthesis generates timbre by adding cosinusoidal oscillations, with k indexing harmonics, r_k denoting amplitude, and φ_k denoting phase.The equation is described as resembling synthesis of a periodic signal with an inverse Fourier series.Source: §2 · pp. 1–2
- In additive synthesis, the McAdams coefficient adjusts harmonic frequencies, changing their distribution and timbre and transforming harmonics into non-harmonic partials or overtones.The coefficient is identified as α.Source: §2 · p. 2
- VTLN warps each speech frame's spectral log amplitude to modify both envelope and pitch, whereas this paper's approach modifies the envelope; pitch changes may be redundant for automatic identification.The passage notes that pitch-modifying approaches may offer greater anonymisation strength when identification is performed by human listeners.Source: §2 · p. 2
3. Anonymisation using the McAdams coefficient
The approach shifts speech formant positions using the McAdams coefficient to alter the spectral envelope. It applies the transformation frame by frame, then randomises the coefficient to produce different pseudovoices and provide some protection from reversibility.
3. Anonymisation using the McAdams coefficient
- The method controls shifts in speech formant positions with the McAdams coefficient, altering the timbre or spectral envelope used in ASV front ends.The paper presents the approach as pseudonymisation through manipulations intended to degrade ASV reliability.Source: §3 · p. 2
3.1. Application to speech signals
- For each frame, LPC analysis yields pole positions; complex poles are shifted by transforming their angles, while real poles remain unchanged, and modified LPC coefficients are combined with the original residual for resynthesis.At a 16 kHz sampling rate, the shift expands or contracts frequencies away from or toward approximately 2.5 kHz; overlap and add combines the processed frames into the anonymised signal.Source: §3.1 · p. 2
3.2. Stochastic anonymisation
- Rather than a fixed α, the stochastic approach samples speaker-dependent coefficients from a uniform range, using different ranges for enrolment and test data to produce different pseudovoices and some protection from reversibility.The original fixed coefficient made anonymisation deterministic and easily reversible; values α > 1 were excluded after listening tests found less acceptable intelligibility degradations.Source: §3.2 · pp. 2–3
4. Experiments
The experiments evaluate McAdams-based anonymisation on VoicePrivacy 2020 data across ASV privacy and ASR intelligibility measures and two adversary scenarios. Randomised variants are competitive in the semi-informed scenario, with privacy gains accompanied by intelligibility costs; retraining and ZEBRA results add further qualifications.
4.1. Data and metrics
- Evaluation uses LibriSpeech and VCTK data, reports average performance despite observed dataset and gender dependencies, and assesses privacy with ASV EER, intelligibility with ASR WER, and privacy evidence with ZEBRA.The system does not use the training partition; the ASV and ASR evaluation systems use LibriSpeech train-clean-360.Source: §4.1 · p. 3
4.2. Scenarios
- The ignorant o-a scenario compares original enrolment with anonymised test speech, whereas semi-informed a-a compares similarly anonymised enrolment and test utterances.The semi-informed adversary knows anonymisation is used and has access to the systems, but not their configuration; the study also reports retrained ASV and ASR results.Source: §4.2 · p. 3
4.3. Results
- In the semi-informed a-a scenario, U(0.7,0.9) reaches 33.8% EER versus B1's 32.3%, while B1 leads in the ignorant o-a scenario.All systems increase EER from the original 3.3%.Source: Figure 4 · p. 4, §4.3 · p. 4
- After ASR retraining, U(0.5,0.9) WER falls from 45% to 9.6%, while EER reaches 37.5%; B1 WER falls from 10.95% to 8.25%.These experiments assess anonymisation robustness when systems are retrained on anonymised data and intelligibility when ASR adapts to it.Source: §4.3 · p. 4
- All systems receive worst-case ZEBRA category C in the semi-informed, retrained-ASV setting, tempering differences in average disclosure.Fewer disclosure bits indicate better privacy, but the worst-case tags are identical.Source: §4.3 · p. 4, Table 1 · p. 4
- 10.95% WER for B1 and 13.7% for U(0.8,0.9), versus 8.45% originally, show the intelligibility cost accompanying McAdams privacy gains.Lower α_min gives better anonymisation but greater WER degradation.Source: Figure 5 · p. 4, §4.3 · p. 4
5. Conclusions
The paper concludes that a randomised McAdams transformation can substantially increase ASV error with modest intelligibility degradation, while offering only partial protection against speaker recognition. It identifies averaging and the limited scope of the transformations as boundaries for further work.
5. Conclusions
- The randomised McAdams transformation increases baseline ASV EER ten-fold with only modest intelligibility degradation, including under a semi-informed adversary.The randomisation also provides some level of protection from reversibility.Source: §5 · p. 4
- True anonymisation remains unmet until error rates reach random performance, and database and gender biases can make average privacy conceal weaker protection for some speakers.The study has so far adjusted formant positions in frequency; future work could explore spectral-envelope pole perturbations, brute-force resilience, and combinations with the primary baseline.Source: §5 · p. 4