📄 The role of context in neural pitch accent detection in English

Summary

The paper addresses automatic detection of pitch accents, the word-level prosodic prominences that often mark new or contrastive information, in English speech. Building on the CNN detector of Stehwien et al. (2018) (SVS18), the authors argue that prominence is relative to a speaker's baseline, so a model needs wider acoustic context. Their speech encoder applies CNN layers to six OpenSMILE prosodic features per frame, pools the frames of each word using corpus timestamps, and passes the whole utterance through a BiLSTM that labels every token. Text-only and speech+text variants are also built. Experiments on the Boston University Radio News Corpus (BURNC) use repeated tenfold cross-validation. Accuracy rises from 87.5% (SVS18) to 88.7%, a state-of-the-art result. A trivial rule that accents every content word scores 82.2%, which matches the text-only BiLSTM, so the authors propose it as the proper baseline. Speech helps mainly where speakers depart from that default. Pitch turns out to be the most informative acoustic cue.

Key points

• Data: the annotated part of BURNC has 5 professional radio speakers, about 2.75 h of read news, 1,721 breath-segmented utterances and 28,489 tokens, of which 15,544 are accented. The majority-class baseline is 54.4%.
• Context ablation (dev set): full-utterance CNN+LSTM reaches 89.1%, against 87.9% for full-utterance CNN-only, 88.6% and 87.3% for three-token inputs, and 85.5% for a single token. The CNN-only model is larger (~14m vs ~12m parameters), so model size does not explain the gain.
• Text-only model: it barely beats the content-word baseline. Shrinking its vocabulary from 3000 to 100 words leaves accuracy unchanged, so it relies mostly on frequency and the function/content-word split.
• Default-violating tokens: on accented function words and unaccented content words, the speech-only model scores 66.7% against 38.2% for text-only. Adding text to speech gives only a small further gain.
• Feature ablation: removing pitch (smoothed F0) hurts most, and voicing features are weakest. A duration-only model, which sees just the frame count per word, reaches roughly the content-word baseline level.

Interesting ideas

• I1. Before crediting a learned model, compare it with the cheapest linguistically informed heuristic (here "accent every content word"), since such a rule can quietly match a neural model.
• I2. Report accuracy separately on the cases where the gold label contradicts the default heuristic, because that is where a model shows real understanding.
• I3. Feed the model a degenerate input that keeps only one trivial signal (constant features, so only word length remains) to measure how much the "rich" input really adds.
• I4. Progressively cap the vocabulary, mapping the rest to UNK, to find out what kind of lexical knowledge a model actually uses.
• I5. An accent on a normally unaccented function word (e.g., "before" when "during" is expected) signals contrast with an implicit alternative.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (learn): Every evaluation could include a one-line rule baseline (e.g., POS-default role assignment) as a negative control, so a component only counts as useful if it beats that rule.
• I2 (learn): Score parsers and frame builders on a "default-breaking" slice, such as non-canonical word order or marked argument roles, to see whether they do more than reproduce the majority pattern.
• I3 (learn): Run witness parsers or the LLM teacher on inputs with lemmas masked or features blanked, exposing how much agreement comes from shallow cues.
• I4 (tokens): Apply vocabulary-capping ablations to the tagger and lemmatizer to check whether they lean on a few hundred frequent function words.
• I5 (frames): Explicit focus markers in text, such as italics, "only", or "even", could similarly trigger a contrast-set alternative in the implicature component, rather than being treated as ordinary modifiers.