Text-to-speech systems that run in real time on a CPU read technical text badly. They pause in the middle of "non-small cell lung carcinoma." They stress the wrong word in "magnetic resonance imaging." They put full emphasis on a term the listener heard two sentences ago. These are phrasing and emphasis errors, not pronunciation errors. Fixing them seems to require knowing the domain. The usual way to supply that knowledge is to place a large pretrained language model in front of the speech generator. That approach helps, but it is far too slow for streaming.
This note proposes a cheaper source of the same knowledge. Statistics on which words co-occur, counted once and offline over a domain corpus such as PubMed abstracts, can be stored as a lookup table. At synthesis time that table can tell a small speech model three things. First, which adjacent words form a single technical term that should not be split. Second, how "central" a term is in the field. Third, whether a term has effectively already been introduced, because the words it most strongly co-occurs with were spoken earlier in the passage. The central empirical question is how much of the prosody benefit of a domain-pretrained language-model front-end this table recovers. We measure that against the language model itself, a small distilled domain encoder, and plain frequency features, at a fixed CPU latency budget.
Core Claim
We claim that, for technical-domain text, most of the phrasing and prominence advantage of a domain language-model text encoder comes from corpus-level collocation structure. That structure is two-sided: certain word sequences behave as one unit, and certain terms are predictable once their neighbours are mentioned. It can be precomputed as a co-occurrence graph and served at constant cost per word.
Concretely, we build a word graph over a domain corpus. Edges are weighted by pointwise mutual information (PMI), a score of how much more often two words appear together than chance predicts:
$$\mathrm{PMI}(a,b)=\log\frac{p(a,b)}{p(a)\,p(b)},$$
Here $p(a,b)$ is the probability that words $a$ and $b$ appear adjacent (for chunking) or in the same sentence (for the graph), and $p(a)$, $p(b)$ are their individual frequencies.
From the graph we derive three per-word features: a term-boundary score from adjacent-word PMI, static topology features (degree, cluster membership, betweenness centrality), and one dynamic feature that we call graph givenness:
$$g(w,t)=\max_{u\in S_{<t}} \mathrm{nPMI}(w,u),$$
Here $S_{<t}$ is the set of content words already synthesized in the current passage, and nPMI is PMI normalized to $[-1,1]$. A high $g$ means the term is predictable from what the listener has already heard. Linguistically, that predicts the term should be de-accented.
These features condition the duration and pitch predictors of a lightweight non-autoregressive synthesizer. The prediction is concrete. On held-out biomedical passages, this system should match a domain-encoder front-end within listener noise on phrase-break accuracy and preference. It should add essentially no real-time-factor overhead. It should clearly beat versions that use only term frequency, or use a co-occurrence graph built from general-domain text.
What Exists, and What It Leaves Open
The motivating evidence comes from work that couples language models to speech generators. Boosting Large Language Model for Speech Synthesis tests several ways to attach LLaMA and OPT to a codec-based speech generator. It finds the best variant is the "coupled" one, which uses the language model only as an upstream text encoder: this gives a 10.9% relative word error rate reduction over vanilla VALL-E. The authors also state that the 7B encoder's latency rules out real-time use, and they train only on general audiobooks.
Separately, the real-time line of work removes heavy text modelling to gain speed. SpeedySpeech removes self-attention in favour of dilated convolutions and still reaches 75.24 on a MUSHRA-style listening test against 62.82 for Tacotron 2. DurIAN replaces attention with an explicit duration model and reports a real-time factor of 0.171 on one CPU core. LPCNet-based TTS runs at 3x real time on one CPU core by handing the vocal-tract filtering to signal processing. All three get their text knowledge from phonemizers and local sequence models. None of them knows that a five-word string is one term.
Domain pre-training work supplies the strong comparison and constrains the design. BioBART continues pre-training BART on 41 GB of PubMed abstracts and shows the objective must respect scientific discourse: sentence permutation hurt. MonoCoder shows a small model pretrained only on the domain beats much larger general models. Durrani et al. argue that domain signal is drowned when mixed into general data. Männistö finds that domain pre-training predicts downstream performance more than the choice of fine-tuning method.
Together these papers tell us that the right comparison is a domain-pretrained encoder, not a general one with adapters. They also tell us our graph should be built on domain text alone. Luan et al. document how technical words take rigid meanings that differ from web usage, which is the reason a general-corpus graph should fail.
The method for the graph comes from bibliometric co-word analysis. The construction (symmetric co-occurrence counts turned into a weighted graph) follows Mohan. The use of betweenness centrality to find bridge concepts follows Yao et al., who identify "MCI" and "Biomarker" as bridges between imaging, genetics, and clinical cores, and Lu et al.. Those studies use the graph only offline, to describe a field. We use it as a runtime lookup.
The known weakness of co-occurrence, that it produces spurious edges, is a direct constraint. We therefore prune low-PMI and low-count edges. We treat the pruning threshold as an experimental variable rather than a fixed choice, and we optionally use the transitivity-based clustering of Clustering Co-occurrence Graph based on Transitivity. The research file found no prior work that connects co-occurrence graphs to speech synthesis prosody.
Research Program
Graph construction. We build the graph once from PubMed abstracts, the same corpus BioBART used, so the graph and the strongest baseline see the same domain text. A matched graph is built from a general English corpus of similar size, and it serves only as an ablation. Words are lemmatized. Adjacent-word PMI gives term-boundary scores, from which we greedily extract multiword terms. Sentence-level co-occurrence gives the weighted graph. Edges are pruned by minimum count and minimum normalized PMI, and the topology features are precomputed per node. The resulting table maps each vocabulary item to a small fixed feature vector plus a sparse neighbour list.
Streaming at runtime. Term-boundary detection needs a bounded lookahead equal to the longest extracted term, which we cap at about six words. Givenness needs only a running set of words already spoken in the passage and a neighbour lookup per new word. Both fit a streaming budget measured in microseconds per word.
The synthesizer. We use SpeedySpeech-style convolutional acoustic models, chosen because they are fast and well characterized. The co-occurrence features are broadcast from words to their phonemes and concatenated into the inputs of the duration and pitch predictors. Speech is generated by a fast vocoder, with BigVGAN as the high-quality reference. The same vocoder is used across all systems so that differences reflect the front-end only. We hold the phonemizer fixed across systems. Rare-term pronunciation is a separate problem, linked to the tokenizer mismatch noted for BioBART, and we keep it out of scope.
Data. The file lists no studio-quality domain TTS corpus, so we must collect one. The acoustic model is trained on LJ Speech for base voice quality. We then record roughly three hours of a single professional speaker reading full PubMed abstracts. Whole abstracts matter, because givenness only appears across sentences. Of these, two hours are used for fine-tuning and one held-out hour for evaluation, and the held-out hour is drawn from topics absent from training. Phrase breaks and prominent words in the held-out hour are annotated by two annotators, and we report agreement. MeQSum questions serve as a second test set of shorter, more informal text. The AfriSpeech-200 clinical split (AfriSpeech-200) provides a clinical ASR model. Following its result that clinical fine-tuning substantially improves recognition, we use that model as an intelligibility probe for synthesized clinical speech.
Baselines
Naive reference point. This is SpeedySpeech fine-tuned on the same domain recordings with phoneme input only, the "do nothing" system. The reported reference numbers on LJ Speech are a MUSHRA score of 75.24 and roughly 5x real time on CPU (SpeedySpeech). Those figures describe general text, and we found no reported result on Lacuna for any lightweight TTS system on biomedical text, so we re-measure everything on our test set. Beating this baseline means significantly higher phrase-break F1 against the human reading, and a lower rate of breaks placed inside annotated terms.
Strongest existing method. This is a domain-coupled encoder: Hao et al.'s Method C (Boosting LLM for Speech Synthesis) applied with the BioBART-Base encoder (BioBART) in place of LLaMA. Its contextual embeddings condition the same SpeedySpeech predictors. We choose BioBART rather than a 7B general model because the domain pre-training evidence says domain beats scale for specialized text. A 7B encoder would also lose on latency before prosody is even measured. The only reported figure for this family is the 10.9% relative WER reduction of coupled LLaMA-7B over VALL-E. We found no reported result on Lacuna for a domain-encoder TTS front-end.
This baseline receives the same fine-tuning data and the same number of training steps. Its encoder may be fine-tuned with or without low-rank adapters, with the better choice selected on a validation split. The proposal "beats" it if it is statistically indistinguishable on phrase-break F1 and listener preference while cutting front-end CPU latency by at least an order of magnitude. A small but significant loss in quality, paired with a large latency gain, would be reported as a trade-off rather than a win.
Small distilled encoder. This baseline is a 4–6 layer convolutional or transformer encoder distilled from BioBART on PubMed text, sized to fit the same CPU budget as our lookup system plus the acoustic model. It is the competitor most likely to win. It gets the same distillation corpus and a hyperparameter search equal in trials to the one we run for graph pruning thresholds.
Ablations isolating the key ingredient. We run frequency and TF-IDF features in place of graph features; graph features built from general-domain text; term-boundary PMI alone, without topology or givenness; and the full system with givenness removed. The givenness ablation is the critical one. It is the only feature that carries cross-sentence context, so it is the clearest test of whether the graph contributes something beyond a multiword-term dictionary.
Why not just…?
Why not just use a frozen general LLM with LoRA as the text encoder? Hao et al. report that their coupled 7B encoder carries latency incompatible with streaming, and that is their own stated limitation, not our inference. The domain-pretraining results from MonoCoder and Männistö further suggest that adapters on a general model underperform a domain-pretrained base. We do not rely on that argument alone: the BioBART-coupled baseline already dominates the general-LLM option on domain fit at far lower cost, and we report its latency directly.
Why not just distill BioBART into a tiny encoder? This is the real threat, and it may win. A distilled encoder sees sentence context, which our static table cannot. The experiment settles it at equal CPU budget. If the distilled encoder matches BioBART and beats the lookup table, the conclusion is that co-occurrence statistics are a weaker substitute for distillation, and we will report that. Our expectation is otherwise, for two reasons. First, a small encoder must spend its capacity storing exactly the term inventory and neighbour structure the table stores for free. Second, the givenness feature uses passage-level history that sentence-level encoders do not receive. The cross-sentence test subset, where terms recur, is where the two should separate. We also test the combination, distilled encoder plus table features, because a gain there would show that the table supplies information the encoder misses.
Why not just fine-tune SpeedySpeech on domain speech and stop? That is our naive baseline, and it measures this alternative directly. The concern is that three hours of speech cannot teach the model the term inventory of a whole field. The held-out topics are chosen so that most multiword terms in the test set never occur in training audio. The term-internal break rate on those unseen terms is the specific number that would show this baseline falling short.
Why not just use TF-IDF or a medical dictionary to mark important words? Frequency cannot say that "patient" is generic while a mid-frequency term bridges two subfields. It also has no notion of givenness. Whether those distinctions actually matter for prosody is an open empirical question, not something the bibliometric papers establish, because those papers never measured speech. The TF-IDF ablation answers it. If TF-IDF matches the full graph, the topology claim is false and the contribution shrinks to term chunking.
Proposed Experiments and Decisive Checks
The primary metrics are phrase-break F1 and prominence F1 against the annotated human reading, and the term-internal break rate on held-out terms. Listening quality is measured with a MUSHRA-style test using the human recording as reference, following the SpeedySpeech protocol so scores are comparable in form. Intelligibility is measured as WER from the AfriSpeech-clinical-adapted recognizer and from a general recognizer. Front-end and end-to-end real-time factor are measured on a single CPU core, matching the DurIAN and LPCNet reporting style.
Vocoder choice is reported but held fixed, informed by the low-latency comparison in Yoneyama et al.. Following LibriTTS, where male voices scored lower naturalness, we would run a second, male speaker if budget allows, so that a gap is not mistaken for a front-end effect.
Three outcomes would be decisive. The first is whether the full lookup system closes most of the gap between the naive baseline and the BioBART-coupled system on phrase-break F1 at roughly the naive system's real-time factor; the headline claim rests on this. The second is whether the general-corpus graph loses most of that gain, which tests the domain claim. The third is whether removing givenness hurts specifically on recurring-term sentences and not elsewhere. That is the one result that would show a static co-occurrence graph supplying context-dependent prosody that a term dictionary and a sentence-level encoder both miss.
We also sweep the PMI pruning threshold. If quality degrades sharply as spurious edges are admitted, that links the known noise problem of co-occurrence mining to a measurable acoustic cost.
References
- Boosting Large Language Model for Speech Synthesis: An Empirical Study
- SpeedySpeech: Efficient Neural Speech Synthesis
- DurIAN: Duration Informed Attention Network for Speech Synthesis
- High Quality, Lightweight and Adaptable TTS Using LPCNet
- BioBART: Pretraining and Evaluation of A Biomedical Generative Language Model
- MonoCoder: Domain-Specific Code Language Model for HPC Codes and Tasks
- A Deep Fusion Model for Domain Adaptation in Phrase-based MT
- A Comparative Study of PEFT Methods for Python Code Generation
- Scientific Information Extraction with Semi-supervised Neural Tagging
- RETAILERS' TECHNOLOGY ADOPTION - A BIBLIOMETRIC REVIEW ON THE WEB OF SCIENCE DATABASE
- Mapping longitudinal scientific progress, collaboration and impact of the Alzheimer’s disease neuroimaging initiative
- Research progress of ferroptosis in Parkinson’s disease: a bibliometric and visual analysis
- Clustering Co-occurrence Graph based on Transitivity
- Clustering Co-occurrence Graph based on Transitivity
- AfriSpeech-200: Pan-African Accented Speech Dataset for Clinical and General Domain ASR
- Comparative Analysis of Fast and High-Fidelity Neural Vocoders for Low-Latency Streaming Synthesis in Resource-Constrained Environments
- LibriTTS: A Corpus Derived from LibriSpeech for Text-to-Speech