💬 NLP

Voice, Bias, and Coreference: An Interpretability Study of Gender in Speech Translation

This study investigates how speech translation models assign grammatical gender to speaker-referring terms by revealing that while internal language models exhibit a masculine bias, high-performing models can override this using acoustic cues and a novel mechanism that links first-person pronouns to gendered terms via distributed frequency spectrum information rather than just pitch.

Original authors: Lina Conti, Dennis Fucci, Marco Gaido, Matteo Negri, Guillaume Wisniewski, Luisa Bentivogli

Published 2026-04-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Lina Conti, Dennis Fucci, Marco Gaido, Matteo Negri, Guillaume Wisniewski, Luisa Bentivogli

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a robot translator that listens to people speaking English and instantly translates their words into languages like Spanish, French, or Italian. In English, you can say "I became a student" without specifying if the speaker is a man or a woman. But in Spanish, French, or Italian, the word for "became" changes depending on the gender: diventato (male) or diventata (female).

The big question this paper asks is: When the robot hears "I became a student," how does it decide whether to use the male or female version? Does it guess based on stereotypes, does it memorize what it saw in its training books, or does it actually listen to the person's voice?

Here is a breakdown of what the researchers found, using some simple analogies.

1. The Robot Doesn't Just Copy-Paste

A common assumption is that if a robot makes a mistake (like calling a woman "he"), it's because it learned that mistake from its training data. It's like a student who just memorized the most common answers in a textbook.

The Finding: The researchers discovered the robot isn't just a copycat. Even when the training data had more examples of the "female" version for a specific word, the robot often still chose the "male" version. It didn't memorize specific word pairs; instead, it learned a general rule of thumb: "In this language, things are usually male unless told otherwise." It's like a chef who assumes every dish is spicy because most dishes in the kitchen are, even if the specific ingredient on the plate is mild.

2. The "Inner Voice" vs. The "Ears"

To understand how the robot thinks, the researchers split its brain into two parts:

The Inner Voice (Internal Language Model): This is the part that knows the rules of the target language (Spanish/French/Italian) but ignores the audio input. It's like a person reading a sentence in a book without knowing who said it.
The Ears (The Encoder): This is the part that actually listens to the voice.

The Finding: The "Inner Voice" is heavily biased toward the male gender. If you ask it to translate without letting it hear the voice, it will almost always guess "male." However, the full robot (with ears) often overrules this inner bias. When it hears the voice, it can say, "Wait, my inner voice says 'male,' but the sound tells me 'female,' so I'll go with female."

3. The Secret Clue: It's Not Just the "Pitch"

For a long time, people thought the robot decided gender by listening to the pitch (how high or low the voice sounds). High pitch = female, low pitch = male. It's like judging a book by its cover color.

The Finding: The researchers used a special "heat map" to see exactly which parts of the sound wave the robot was looking at. Surprisingly, the robot wasn't focusing mostly on the pitch. Instead, it was paying attention to the formants.

The Analogy: Think of pitch as the volume knob, but formants as the timbre or the unique "color" of the voice (like the difference between a cello and a violin playing the same note). The robot was looking at the complex shape of the sound waves (specifically the first and second formants) rather than just the high/low note. It's like identifying a person not just by how loud they shout, but by the unique texture of their voice.

4. The "I" Connection

The most interesting discovery is how the robot connects the voice to the gendered word.
When a speaker says, "I became a student," the word "I" doesn't have a gender in English. But the robot realized that the word "I" is the bridge to the speaker's voice.

The Finding: The robot uses the word "I" (and other self-referential words like "me" or "my") as an anchor. It links the acoustic clues in the voice (the formants) directly to the word "I." Once it knows the voice belongs to the "I," it applies that gender to the rest of the sentence.

The Analogy: Imagine a detective trying to identify a suspect. The suspect says, "I did it." The detective doesn't just look at the word "I"; they look at the voice saying "I." Once the voice is identified as female, the detective knows the whole sentence is about a female. The robot does the same thing: it uses the pronoun "I" to grab the gender information hidden in the voice and apply it to the translation.

5. Why Some Robots Are Better Than Others

The study compared two types of robot architectures (Transformers and Conformers).

The Transformer (The Better Detective): This model was very good at using the voice clues to override its "male default" bias. It relied heavily on the "I" connection to get the gender right.
The Conformer (The Struggling Detective): This model was less accurate. It tended to stick more to its "Inner Voice" (the male bias) and didn't listen to the voice clues as effectively.

Summary

The paper concludes that speech translation robots are smarter than we thought, but they have a specific way of working:

They have a built-in bias to guess "male" by default.
They don't just memorize training data; they learn general patterns.
They don't rely solely on the "high/low" pitch of a voice.
Instead, they use the word "I" as a hook to grab the subtle, complex acoustic clues in the voice (the formants) to figure out the speaker's gender and translate correctly.

If we want to fix robots that get gender wrong, we can't just change the training books or tweak the pitch; we need to understand how they are linking the word "I" to the sound of the voice.

1. Problem Statement

Speech Translation (ST) systems face unique gender bias challenges distinct from text-based Machine Translation (MT). While text systems rely on linguistic context, ST systems process audio that contains acoustic cues (e.g., pitch, formants) revealing the speaker's sociodemographic attributes, including gender.

The Core Issue: When translating from gender-neutral languages (like English) to languages with grammatical gender (e.g., Spanish, French, Italian), ST models must assign gender to speaker-referring terms (e.g., translating "I have become a student" to diventato vs. diventata).
The Risk: Models may misgender speakers by relying on:
1. Training Data Imbalances: Defaulting to masculine forms due to skewed corpora.
2. Acoustic Assumptions: Incorrectly inferring gender based on vocal characteristics (e.g., assuming low pitch equals male).
The Gap: It is poorly understood how ST models utilize acoustic information to make these decisions or whether they rely on specific acoustic features (like pitch) or linguistic coreference.

2. Methodology

The authors employ a multi-faceted interpretability approach to dissect the decision-making process of ST models across three language pairs (English $\to$ Spanish, French, Italian).

A. Datasets and Models

Data: MuST-SHE, a benchmark containing English speech with gender-neutral terms requiring gendered translation, annotated with the speaker's self-identified gender.
Models: Two architectures trained on the MuST-C corpus (to ensure transparent training data analysis):
1. Transformer (Wang et al., 2020): Multilingual encoder-decoder.
2. Conformer (Papi et al., 2024): Monolingual Conformer encoder with Transformer decoder.

B. Analytical Techniques

Training Data Prevalence Analysis:
- Compares the frequency of gendered terms in the training corpus against the model's generation preferences.
- Goal: Determine if models simply memorize term-specific training statistics or learn broader patterns.
Internal Language Model (ILM) Approximation:
- Replaces the encoder's audio output with a zero vector, isolating the decoder's behavior.
- Goal: Measure the decoder's inherent bias (learned language patterns) independent of the input audio.
Contrastive Feature Attribution:
- Uses Conti et al. (2025)'s method to generate saliency maps on input spectrograms.
- Mechanism: Systematically masks segments of the spectrogram to observe changes in the probability of generating a masculine vs. feminine form.
- Validation: Identifies "flipped" examples where occluding specific features changes the gender prediction, establishing a causal link between input features and output.

3. Key Results

A. Training Data vs. Model Behavior

Finding: Models do not simply replicate term-specific gender associations from the training data.
Evidence: While the training data shows a masculine skew (approx. 68–71% prevalence), models frequently predict the less prevalent gender for specific terms (e.g., 87.8% of feminine predictions in Spanish were for terms where the masculine form was more frequent in training).
Conclusion: Models learn a broader "masculine prevalence" pattern rather than memorizing individual word pairs.

B. The Role of the Internal Language Model (ILM)

Finding: The decoder (ILM) exhibits a strong masculine bias (preference scores of 0.74–0.81 for the Transformer), often stronger than the raw training data prevalence.
Interaction:
- The Transformer model frequently overrides this internal bias when acoustic input is present, successfully generating feminine forms despite the ILM's preference for masculine.
- The Conformer model relies more heavily on its ILM and struggles to override the bias using acoustic cues, resulting in lower gender accuracy.

C. Acoustic Cues: Pitch vs. Formants

Finding: Contrary to the assumption that pitch (Fundamental Frequency, F0) is the primary driver for gender assignment, the models rely more heavily on formants (F1 and F2).
Evidence: Saliency maps show the highest attribution scores in the 350–2,500 Hz range (formants), which are crucial for vowel identity and vary by speaker gender. The pitch range (80–350 Hz) is less dominant, though still present.
Implication: Interventions that only manipulate pitch are insufficient to neutralize gender bias in ST.

D. The Coreference Mechanism

Finding: The most critical linguistic feature for gender assignment is the first-person pronoun "I" (and self-referential words like "my," "me").
Mechanism:
- In text-based MT, models use gendered pronouns (he/she) to resolve ambiguity.
- In ST, the word "I" is semantically gender-neutral. However, the model uses "I" as an anchor to link the acoustic cues of the speaker's voice to the grammatical gender of the target term.
- Statistical Proof: For the Transformer model, "I" is the top-scoring word in 36–46% of "flipped" examples (where occlusion changes the prediction). Including all self-referential words raises this to 43–53%.
- Asymmetry: This mechanism is significantly more critical for generating feminine forms (60%+ flip rate when occluded) than masculine forms, suggesting the model uses acoustic cues to overcome its default masculine bias specifically for feminine outputs.

4. Key Contributions

Decoupling Bias Sources: Demonstrates that ST gender bias is not merely a reflection of training data distribution but a complex interplay between a biased internal language model and acoustic input.
Acoustic Feature Discovery: Identifies that formants (F1/F2), not just pitch, are the primary acoustic drivers for gender assignment in ST models.
Coreference in Audio: Reveals a novel mechanism where models use semantically neutral first-person pronouns ("I") as a bridge to access distributed acoustic gender information, effectively performing "acoustic coreference resolution."
Architectural Comparison: Shows that Transformer-based ST models are better at leveraging these acoustic cues to override internal biases compared to Conformer-based models.

5. Significance and Implications

Mitigation Strategies: Current mitigation strategies focusing solely on rebalancing training data or manipulating pitch are likely insufficient. Effective interventions must address the model's reliance on formants and its tendency to default to masculine forms when acoustic cues are ambiguous or missing.
Ethical Considerations: The study highlights the risk of "misgendering" speakers whose vocal characteristics do not align with binary gender expectations. Understanding that models use "I" + voice to infer gender suggests that systems could be designed to respect user-specified pronouns rather than inferring gender from voice.
Future Research: The findings suggest that future ST systems (including SpeechLLMs) should be evaluated on their ability to handle non-binary gender identities and their reliance on specific acoustic features beyond pitch.

In summary, the paper provides a granular, causal explanation of how speech translation models navigate gender ambiguity, revealing that they function by anchoring acoustic voice characteristics to first-person linguistic markers, rather than relying on simple pitch detection or training data memorization.