Analysis, synthesis, enhancement, and classification of audio and speech.
A human-inspired framework combines acoustic evidence with linguistic knowledge, reporting state-of-the-art phonetic error rates from 100 hours of training data and zero-shot results in 95 unseen languages.
Restricting CTC alignment states to batch target tokens preserves full-vocabulary loss and gradients while cutting full-step memory 5.1× at 180K tokens.
WenetSpeech-Min pairs Minnan and Mandarin text for around 10,000 hours of online speech, creating shared training and evaluation material for ASR and TTS.
A shared autoregressive–non-autoregressive model writes lead sheets before audio tokens, raising expert overall-quality preferences to 49.3% versus 34.6% without planning.
BEACON ensembles word timestamps from off-the-shelf ASR systems to curate 413 hours of CHILDES audio, reducing average out-of-domain WER by up to 19.5%.
An ABC notation plan edited by a text LLM raises SongBench’s mean score from 6.6438 to 6.7220 before token-level audio rendering.
A parallel FOA encoder adds spatial latents to Omni LLMs, raising Qwen3-Omni's MMAU-Pro spatial score by 9.50 points while retaining general-audio performance.
A curated 830-hour dialogue dataset couples style-conditioned instructions with synthesized speech, raising multi-dimensional control scores without reported losses in conversational capability.
VibeVoice-ASR-Streaming interleaves audio, lookahead, and prior text so one model can produce speaker-attributed transcripts as speech arrives.
A 30-hour triple-annotated benchmark separates end-of-turn and interruption decisions, showing that low-latency interruption detection still produces excessive false positives.
