Translating a song so it can still be sung means fitting the new words to the old melody: roughly one syllable per note, stressed syllables on strong beats, and no word split across a rest. In languages with grammatical gender, the feminine form of a word is often longer than the masculine one. Italian dottore has three syllables and dottoressa has four. Spanish el doctor has three and la doctora has four. A translation system that must hit exactly eight syllables therefore has a cheap way to satisfy the melody: pick the masculine form. We hypothesize that singability constraints actively manufacture masculine default bias, which is the tendency of translation systems to output masculine forms even when the source says otherwise. If so, the current way of stating those constraints is the cause: a single target syllable count per line, typed in by hand.
The proposal has three parts. First, we measure this "melodic gender penalty" directly. Second, we test a controlled prediction: the penalty should appear only where feminine and masculine forms differ in length. Third, we test a specific fix. Instead of a scalar syllable count, the decoder receives a note-level grid extracted automatically from a sung performance by forced alignment, meaning software that time-stamps each sung syllable in the audio. The grid records note durations, rests, and metric strength. It shows where the melody has legal slack, such as a long note that can carry two syllables or two short notes that can share one. With that information, the decoder can make room for a feminine ending rather than dropping it. The claim is falsifiable at both ends. If no penalty exists, the intervention is moot. If a simple tolerance of plus or minus one syllable removes the penalty, the grid is unnecessary.
Core Claim
Singable translation systems enforce length as one hard number per line. Songs Across Borders prompts mBART with a target syllable count, rhyme, and mandatory word boundaries, and reports 99.85% length accuracy and over 95% boundary recall. Masculine defaults are already severe without any length pressure. Investigating Failures of Automatic Translation in the Case of Unambiguous Gender finds that M2M-100 stays below 70% accuracy even when gender is stated explicitly. Evaluating Gender Bias in Machine Translation reports 80% versus 54% French accuracy on stereotypical versus non-stereotypical roles.
Nobody has asked whether a near-perfect length constraint buys its accuracy partly by sacrificing feminine forms. We define the penalty as a difference-in-differences, so that ordinary bias is separated from bias caused by the constraint:
$$P = \big(A_F^{\text{free}} - A_F^{\text{con}}\big) - \big(A_M^{\text{free}} - A_M^{\text{con}}\big)$$
Here $A_F$ and $A_M$ are gender accuracy on feminine-target and masculine-target items. "Free" means unconstrained translation and "con" means translation under the melody constraint. $P>0$ means the constraint hurts feminine targets more than masculine ones.
The central prediction is a dose-response. $P$ should grow with $\Delta s$, the extra syllables in the feminine form relative to the masculine one. $P$ should be near zero for items where the two forms have equal length, such as Spanish el cantante and la cantante, where only the article changes and the article has the same syllable count. This length-matched control is what turns a correlation into a mechanism.
Research Program
Benchmark construction. No existing singable-translation resource covers gendered target languages. The 162-song set in A Computational Evaluation Framework for Singable Lyric Translation targets Japanese and Korean, which have no grammatical gender. We will therefore build an English to Italian, Spanish, and French set. It will contain lyric lines that carry an unambiguous gender trigger, either in the third person ("the doctor sang to her daughter, she…") or in the first person with a stated persona, which mirrors the speaker-gender category of MuST-SHE. Triggers are inserted with the phrase-structure templating approach of Investigating Failures of Automatic Translation in the Case of Unambiguous Gender, constrained so that each line keeps the source melody's syllable count. Each item is tagged with $\Delta s$ and with whether the gendered word falls on a strong or a weak beat.
Gender is evaluated along whole agreement chains: article, noun, adjective, and participle, following Under the Morphosyntactic Lens. That paper shows that once the noun's gender is chosen, the rest of the chain usually agrees with it, which tells us to score chains rather than isolated nouns. Gender is supplied as an explicit input, never inferred from the singer's voice. A singer's voice is not the persona's gender, and inferring identity from the voice is ethically fraught.
Grids from performances. Each line is recorded or sourced as an English sung performance. We align it to its lyrics with phoneme-level aligners: the Montreal Forced Aligner, the two-pass approach of WhisperX, and the Qwen3-ASR aligner. The aligner output is combined with pitch tracking, following Improving Lyrics Alignment through Joint Pitch Detection, to recover note onsets, held notes, melismas, and rests.
Sung vowels are long and pitch-modulated, which makes singing much harder to align than speech (Semi-supervised Lyrics and Solo-singing Alignment; Acoustic Modeling for Automatic Lyrics-to-Audio Alignment). For that reason we keep a subset with symbolic scores, so the extracted grid can be scored against ground truth before anything downstream depends on it. This step directly addresses the manual-prompt limitation that Songs Across Borders acknowledges.
Grid-constrained decoding. The decoder does not receive "8 syllables." It receives a sequence of note slots, each with a duration, a metric strength, and a flag for whether it is followed by a rest. Syllables may map one-to-one onto slots. A syllable may also take two slots, the many-to-one mapping that SongGLM encodes for melismas. Two syllables may share one long slot, which matches the vowel merging (synalepha) that is standard in Romance-language singing.
Decoding is scored with the stress-duration rule of Yufei Tian, which places stressed syllables on long notes, together with the beat-stress correlations documented by Dan Morris et al.. The rule matters here because feminine suffixes can move stress, as in dottóre versus dottoréssa. The gender of each agreement chain is a hard constraint. Length is a soft cost spread across the whole line, so the decoder can pay for an extra feminine syllable by dropping a filler word or using a melisma elsewhere.
Tokenization as a measured confounder, not a premise. Syllable tracking during decoding runs over subword pieces, the word fragments a byte-pair-encoding tokenizer produces. Feminine words are split into more pieces (Exploring the Impact of Training Data Distribution and Subword Tokenization on Gender Bias), so the syllable estimate for a partially generated word may be less accurate for feminine forms. That would push the decoder away from them. The same paper, however, finds that the link between subword count and accuracy disappears once training frequency is controlled. Under the Morphosyntactic Lens reports instead that character-level models handle feminine forms better than BPE models.
We do not take a side in advance. We will measure the decoder's syllable-estimation error on feminine versus masculine continuations with word frequency held fixed. We will then swap in a character-level syllable counter, which counts syllables from the characters of the partial word and has no piece-level ambiguity, to see whether any remaining penalty is caused by tokenization.
Baselines
The naive reference point is unconstrained translation with the same mBART50 backbone and with a strong instruction-tuned LLM. It gives $A^{\text{free}}$, the masculine-default floor before any melody pressure.
The strongest existing singability method is Songs Across Borders, retrained with its released recipe on the same back-translated data for each target language. It is given the gold scalar syllable count and boundaries, and the gender is supplied by an explicit tag. It reports 99.85% length accuracy and over 95% word-boundary recall. No reported result was found on Lacuna for its gender accuracy, which is precisely what we measure.
The strongest gender-side reference is the "unambiguous trigger" setting, in which M2M-100 is reported below 70% accuracy (Investigating Failures). For audio-conditioned gender cues, MuST-SHE reports an end-to-end difference score of 39.5 against 28.0 for cascades (Gender in Danger?). Because gender is given as input here, these numbers serve as context rather than direct targets.
All systems share the backbone, training data, decoding beam width, and a matched hyperparameter search budget. The method counts as beating the baselines if it reduces $P$ by at least half relative to the Songs Across Borders baseline. It must do so with note-fit and boundary recall within one point of that baseline, and with human-judged singability that is not worse.
The ablations isolate one ingredient each. The first is the full grid with gold symbolic notes, which separates alignment noise from the grid idea. The second is a scalar budget with a tolerance of plus or minus one syllable, which is the cheap alternative. The third is the grid with the stress term removed. The fourth is the character-level syllable counter.
Why not just…?
Why not just prompt an LLM with "female singer, 8 syllables"? This is a real threat, and it is a baseline rather than a straw man. The evidence against it is indirect. Explicit gender statements already fail to override stereotyped priors (Investigating Failures). Feature attribution shows that models give very little weight to gender-marking tokens when they default to masculine (Attanasio). If the prompted LLM shows $P\approx 0$ and also matches the grid system on syllable fit, the central claim is wrong for modern LLMs, and we will report that.
Why not just allow a one-syllable tolerance? Singers do absorb extra syllables, so this may be enough. Its weakness is that it does not say where the extra syllable goes. An unplaced syllable can land on a weak beat or straddle a rest, which are the failure modes that Songs Across Borders and the stress-duration work identify. The tolerance ablation tests this directly. If it closes the penalty with equal stress alignment and singability ratings, we drop the grid and recommend the tolerance.
Why not just generate n-best translations and filter for correct gender? If the constrained beam rarely contains a correct feminine candidate that fits the length, filtering has nothing to choose from. We will measure how often a gender-correct, length-valid candidate appears in the top-k as a function of $\Delta s$. High coverage would make reranking a sufficient and cheaper fix.
Why not just fix the tokenizer, for example by adding feminine forms as whole tokens? Tokenization Matters reports that vocabulary parity for neopronouns raised consistency from 14.1% to 58.4%. In contrast, Exploring the Impact… found that preventing word splits brought little benefit and could hurt general performance, while embedding fine-tuning moved German $\Delta G$ from 9.6 to -2.9. Given this conflict, tokenization enters only as a frequency-controlled diagnostic and the character-counter ablation.
Decisive Checks
The first check is existence. Under the Songs Across Borders baseline, $P$ must be significantly positive. The second is mechanism. $P$ must rise with $\Delta s$ and vanish on length-matched items. If $P$ is flat across $\Delta s$, the bias is not caused by melody pressure and the project ends at the measurement paper. The third is the fix. With gold notes, the grid decoder must cut $P$ without losing singability. With aligner-derived notes, it must retain most of that gain, and comparing the two quantifies the cost of automatic alignment on sung audio. Singability is judged by trained singers who perform the translated lines blind to the system that produced them. Automatic scores accompany these judgments: syllable distance and musical-structure distance from A Computational Evaluation Framework for Singable Lyric Translation, plus the stressed-on-strong-beat rate.
References
- Songs Across Borders: Singable and Controllable Neural Lyric Translation
- Investigating Failures of Automatic Translationin the Case of Unambiguous Gender
- Evaluating Gender Bias in Machine Translation
- A Computational Evaluation Framework for Singable Lyric Translation
- Under the Morphosyntactic Lens: A Multifaceted Evaluation of Gender Bias in Speech Translation
- Montreal Forced Aligner: Trainable Text-Speech Alignment Using Kaldi
- WhisperX: Time-Accurate Speech Transcription of Long-Form Audio
- Qwen3-ASR Technical Report
- Improving Lyrics Alignment Through Joint Pitch Detection
- Semi-supervised Lyrics and Solo-singing Alignment
- Acoustic Modeling for Automatic Lyrics-to-Audio Alignment
- SongGLM: Lyric-to-Melody Generation with 2D Alignment Encoding and Multi-Task Pre-Training
- Unsupervised Melody-to-Lyrics Generation
- Relationships Between Lyrics and Melody in Popular Music
- Exploring the Impact of Training Data Distribution and Subword Tokenization on Gender Bias in Machine Translation
- Gender in Danger? Evaluating Speech Translation Technology on the MuST-SHE Corpus
- A Tale of Pronouns: Interpretability Informs Gender Bias Mitigation for Fairer Instruction-Tuned Machine Translation
- Tokenization Matters: Navigating Data-Scarce Tokenization for Gender Inclusive Language Technologies