LASE: Fixing Accent Bias in AI Speech Recognition with Language-Adversarial Speaker Encoding

Paper: LASE: Language-Adversarial Speaker Encoding for Indic Cross-Script Identity Preservation Author: Venkata Pushpak Teja Menta arXiv: 2605.00777 | 2026-04-30

1. The AI That Judges You by Your Accent

Imagine a speech recognition system where the same person saying "Hello" in English versus in an Indian language is treated as two different people. Not because the voice truly changed — but because the system learned to factor *language/accent* into *speaker identity*.

This is a hidden bias in today's speaker encoders: they treat the language being spoken as part of who the speaker is.

2. The Identity Crisis in Cross-Script Voice Cloning

A key requirement in multilingual voice cloning:

> The same speaker should be recognized as the same person, no matter which language or script they speak.

But existing encoders fail:

• WavLM-base-plus-sv: similarity drops by 0.082 when the same voice switches script
• ECAPA-TDNN: drops by 0.105

Worse, the bias is accent-conditioned:

• Western accents (speaking English, Hindi, Telugu, Tamil): large gaps
• Indian accents: small gaps (WavLM drops only 0.006)

This means the system discriminates more severely against "non-standard" accents.

3. LASE: Language-Adversarial Speaker Encoding

The paper proposes LASE (Language-Adversarial Speaker Encoding).

Core idea:

> Train the encoder to "forget" language information and keep only speaker identity.

Technical approach:

• Adversarial training: the encoder tries to make a discriminator unable to guess which language is spoken
• While simultaneously ensuring speaker identity remains correctly recognizable
• Result: language information and speaker information are disentangled

Think of it as training an interviewer: don't judge who the candidate is by their accent — judge them by their unique voiceprint.

4. Why It Matters

Applications of cross-lingual speaker recognition:

• Multilingual voice assistants: a user interacting in different languages should be recognized as one person
• Cross-lingual TTS: generate speech in many languages with one person's voice
• Voice biometrics: security systems shouldn't fail when users switch languages
• Speech translation: preserving speaker identity consistency

Most importantly, fairness:

• Current systems show systematic bias against non-Western accents
• LASE provides a technical approach to reducing this bias

5. Feynman-Style Judgment: Correlation vs. Causation

Feynman stressed distinguishing correlation from causation:

> "Just because two things happen together doesn't mean one causes the other."

In speech recognition:

> "Language (correlation) should not be confused with speaker identity (causation). The same person speaking different languages is still the same person."

Current encoders err by binding language and identity — because they co-occur in training data, the model wrongly assumes they are the same thing. LASE uses adversarial training to force the model to decouple these signals — a direct application of scientific method in AI.

6. Takeaways

If you're building multilingual AI systems, ask:

1. "Does my system treat 'language' as part of 'identity'?" 2. "Are there hidden biases in cross-lingual scenarios I haven't noticed?" 3. "Could adversarial training help disentangle features that shouldn't be conflated?" 4. "Is my system fair to non-standard accents and dialects?"

LASE reminds us: AI bias often comes from spurious correlations in training data.

Just because "English speakers" and some acoustic feature co-occur in the data, the model binds them together. Breaking that binding requires adversarial training objectives — forcing the model to learn the features that *truly matter* while ignoring the distractors.

In AI, learning to *ignore* is just as important as learning to *attend*.