NAVER LABS Europe teaches LLMs to write text that sounds right when spoken
Researchers at NAVER LABS Europe have studied how to make large language models produce text that works well when read aloud by a Text-to-Speech system, rather than fixing it afterwards. In their EMNLP 2026 paper, they report that the Feature-aware Sampling and Tuning (FaST) framework gave the best balance between TTS-friendliness and helpfulness.
- NAVER LABS Europe studies how to make LLMs natively generate text suited to Text-to-Speech.
- The authors introduce two preference datasets, CORA and Recipe, with paired TTS-friendly and TTS-unfriendly responses.
- The evaluation suite combines a heuristic metric, a TTS to ASR pipeline and a MUSHRA listening study.
- According to the paper, FaST achieves the best overall tradeoff between TTS-friendliness and helpfulness.
- The metrics correlate strongly, so an efficient heuristic can reliably assess TTS-friendliness.
The problem: text written for reading, not listening
The paper, titled "Ready to speak: aligning LLMs for TTS-friendly text generation", is by Thibaut Thonet, Jos Rozen and Laurent Besacier. It is listed for the Conference on Empirical Methods in Natural Language Processing (EMNLP) in Budapest, Hungary, running from 24 to 29 October 2026.
According to the authors, current Large Language Models are primarily optimized for written text. Their outputs are often grammatically correct and helpful, but poorly suited for spoken delivery through Text-to-Speech (TTS). That matters for any voice interface that reads a model's answer aloud.
Alignment instead of rewriting
The team frames the task as a preference alignment problem. Instead of relying on downstream rewriting modules that convert a written answer into a spoken-style one, they directly align LLMs so that the generated text is optimized for spoken delivery from the start.
To support this, the authors introduce two preference datasets covering different target domains, called CORA and Recipe. Each contains paired TTS-friendly and TTS-unfriendly responses, which can be used to teach a model which style to prefer.
A three-part evaluation suite
The researchers also propose an evaluation suite with three components:
- a pattern-based heuristic metric;
- a TTS to ASR evaluation pipeline;
- a MUSHRA listening study with human judges.
The authors report a strong correlation between these different metrics. In their view, this shows that TTS-friendliness can be reliably assessed with an efficient heuristic, without always running a full listening study.
FaST compared with alignment baselines
In the experiments, the team compares the recently proposed Feature-aware Sampling and Tuning (FaST) framework against an array of alignment baselines. According to the paper, FaST leverages interpretable features instead of a black-box reward model.
The authors write that they "found that FaST achieves the best overall tradeoff between TTS-friendliness and helpfulness across various settings." The source does not give numerical results in this summary.
What it means
The work points to a way of building voice-ready language models by changing how they are trained, rather than adding a separate rewriting step. The released datasets and the heuristic metric could give other researchers a common basis for testing spoken-style generation, and the reported agreement between metrics suggests that cheaper automatic checks can stand in for costly listening tests.
NAVER LABS Europe: Ready to speak: aligning LLMs for TTS-friendly text generation
