TTS Control, Smarter Voice Agents, and Identity-Preserving Video

Word-level emphasis conveys meaning that text alone cannot express and is therefore essential for a practical text-to-speech (TTS) system. In this work, we study how to add emphasis control post hoc to a pretrained TTS model that has never been trained with emphasis annotations. We present EmpDPO, a simple, efficient, and effective two-stage recipe that adds emphasis control to a multilingual zero-shot diffusion TTS model through supervised fine-tuning followed by direct preference optimization (DPO). The training signals for both stages come from an off-the-shelf multilingual emphasis detector, so that no manual emphasis annotation or task-specific reward design is needed. We find that DPO can substantially strengthen emphasis control with only a few thousand preference pairs and training steps. Objective and subjective evaluations against strong public systems show that our model produces perceptible emphasis while preserving naturalness and overall quality.

Zero-shot text-to-speech (TTS) models clone a target voice from a short reference clip, commonly referred to as the prompt. In practice, the prompt has to be cut from a longer recording of the speaker, so which part to use must be chosen. To understand what makes a good prompt, we analyze hundreds of thousands of outputs synthesized by four TTS models from candidate prompts of YouTube recordings. The choice of prompt substantially changes intelligibility, naturalness and speaker similarity. Standard predicted-quality scores capture mainly the naturalness part but miss most of the intelligibility gain. We find that the models largely agree on which prompt yields the highest measured speaker similarity, but much less on which prompt yields intelligible speech. We propose a small network that learns from the outputs a TTS model produced to rank a recording's candidate prompts from their audio alone. Without synthesizing anything at selection time, it picks prompts that lower the word error rate on held-out recordings by a third (Qwen3-TTS) to two fifths (F5-TTS) relative to a randomly chosen prompt, more than predicted quality achieves and nearly as much as synthesizing a dozen test sentences from every candidate first.

Voice-cloning TTS models synthesize an unseen speaker's voice from brief speech prompts, but it remains unclear which acoustic properties they represent and use. We introduce a feature-based interpretability framework for autoregressive voice-cloning TTS models, analyzing features from sparse autoencoders (SAEs), PCA, and random directions via sparse probing, steering, and quantitative auto-interpretability. We focus on Qwen3-TTS and extend probing and steering to $5$ additional models to test generalization across architectures. These models encode concepts like pitch, energy, noise, and accent, but sparse prediction and steering are effective only for some concepts. Strong probing does not always translate to steering or vice-versa. SAEs are useful for sparse probing of human-labelled concepts and novel-concept discovery, while PCA is competitive for dense probing and steering. Auto-interpretability finds $684$ automatically validated SAE features in Qwen3-TTS covering speaker clusters, speaking style, and recording conditions beyond known concepts. Overall, we perform a comprehensive interpretability study of autoregressive voice-cloning TTS models.

Speech technology for Irish must be built from very little transcribed speech, yet the language's pronunciation is thoroughly documented in dictionaries and dialect descriptions. We ask whether that documentation can supervise the adaptation of a large pretrained text-to-speech model to a language it was never trained on. We adapt Qwen3-TTS to Irish using 12.8 hours of crowdsourced Common Voice speech, correcting two defects in the vendor fine-tuning script that training loss does not reveal (Appendix A), and reach character error rate (CER) 0.225 on held-out cross-corpus text, against 0.397 for an exhaustively tuned from-scratch system and 0.437 for zero-shot cloning. We then compile the Wiktionary pronunciation record into a lexicon and a phoneme critic, and use that critic as the verifier in one round of rejection fine-tuning: CER 0.159 cross-corpus and 0.067 in-domain, with the voice still anonymous across all 202 training speakers. The same critic explains why the voice cannot yet be steered to a minority dialect: 229 of 12,311 training clips lean Ulster against 8,451 Munster. Pronunciation documentation, converted into a verifier, is a practical lever for low-resource speech synthesis, and a way to measure what the available data cannot support.

Semantic completeness alone cannot determine when an assistant should respond in a multi-speaker environment: a completed utterance may belong to an interfering speaker, while the designated user is still speaking, acknowledging, or requesting time. We therefore study semantic turn inference conditioned on speaker identity. We propose SpeakerTurn, a framework that jointly predicts the target transcript and turn state from a speech mixture and an independent target-speaker reference. Built on CocktailASR-1, it links semantic decisions to the corresponding target content through joint supervision, distinguishing complete, incomplete, backchannel, wait, and target-absent states, with a separate interruption task. We also examine whether structured chain-of-thought (CoT) supervision improves decisions by representing speaker presence and semantic cues. On a synthetic multi-speaker test set, SpeakerTurn achieves five-state macro F1 of 0.811 and target-transcription character error rate (CER) of 4.20%, correctly resolving all ten reference-swapped interruption pairs. Joint-format transcription and native-interface recognition nevertheless behave differently, and CoT improves validation transcription without improving turn macro F1. These results show why target attribution, semantic turn discrimination, and recognition retention require separate assessment when adapting speech language models for speaker-conditioned interaction.

End-of-turn (EoT) detection has been studied mostly through features, models, and modalities, while the speech genre of the training data has drawn far less attention. We isolate this factor. Using one reproducible annotation pipeline, we build a matched pair of corpora, multi-turn human-human conversational and single-turn human-machine, and train the same streaming ConvLSTM EoT detector on each. Across four test sets the conversational model reaches up to +0.45 higher F1, driven by fewer false alarms on the mid-turn pauses and hesitations that are rare in single-turn corpora. This conversational-training effect is not specific to frame-level decoding. Under a matched complete-or-not test it also appears at the utterance level, across two encoders, but only on conversational test data. Testing only on single-turn speech, as most prior work does, therefore hides the benefit of conversational training data. Against open-source baselines, our 1.8M-parameter model attains the highest F1 on every test set.

We present a cascade streaming voice privacy system, consisting of ASR, NLP, and TTS modules. By leveraging recent improvements in all three aspects of speech and language processing, we show that such a system is capable of achieving a result comparable to state-of-the-art offline voice privacy systems, while providing controllable capabilities of content protection beyond voice. With an estimated utterance-onset end-to-end latency of 717 ms, the presented system preserves speech content with 11% WER on Librispeech, while detecting PII with a 69% SLU F1 score and achieving up to 47.66% EER against a semi-informed attacker.

Severe blur, downsampling, noise, and compression can erase the identity cues needed to faithfully restore a person's face. We therefore use photographs of the same person as an external source of identity. AlbumVSR conditions a frozen, one-step streaming diffusion transformer on a personal album. We use four MagFace/CLIP set tokens to encode the album, and 64 texture tokens which come from a fused UV atlas rendered at the current window's estimated face pose. These tokens are appended to the transformer's cached cross-attention context. On an identity-disjoint VoxCeleb2 benchmark with heavy synthetic degradation, AlbumVSR raises ArcFace similarity to the ground truth from 0.586 to 0.670 and face PSNR from 24.98 to 26.71 dB. On the 91 test identities, its mean identity similarity is higher than that of all 14 baselines.

Spoken Dialogue Models (SDMs) typically share a speech-token space for generation and synthesis, leading to long autoregressive sequences and costly inference. Lower-rate speech tokenizers shorten sequences but may introduce compression-quality trade-offs and costly replacement without breaking this coupling. To address this limitation, we propose CoSTA, a compatibility-preserving post-hoc framework that decouples generation and synthesis token spaces while retaining the original speech tokenizer and decoder. CoSTA extends the generation token space with compressed representations via mean-initialized vocabulary extension, pronunciation adaptation, and dialogue adaptation, followed by deterministic expansion back to the original speech-token space before synthesis. On bilingual speech-dialogue benchmarks, CoSTA achieves approximately 1.7times expanded speech-token generation speed, lower full-response latency, and higher throughput while preserving response quality. Analyses show more sustained interaction under limited budgets and a practical trade-off among compression, decoding efficiency, and quality. Further experiments demonstrate cross-architecture transfer and complementarity to low-rate tokenization.