Flash RNN‑T is a new implementation for the recurrent neural network transducer that combines specialized kernels with optional tiled recomputation to reduce memory usage while maintaining high throughput. On NVIDIA H100 GPUs, the untiled version delivers 24–39% higher throughput and 26–56% less memory than the fastest baseline, and the tiled version can use nine times less memory on medium workloads with only a 16% throughput drop. For long workloads, tiled Flash fits within 80 GB of memory where all other tested alternatives run out of memory, achieving at least 95% of peak throughput in its lowest‑memory configuration.
arXiv:2601. 06199v3 Announce Type: replace-cross Abstract: Scaling Multimodal Large Language Models (MLLMs) to long-form speech is bottlenecked by the explosive growth of input tokens.
arXiv:2607. 20086v1 Announce Type: cross Abstract: State-space sequence models are attractive for streaming speech because they maintain compact recurrent state, but scan-style training kernels can have unfavorable constants for short audio tasks.
arXiv:2607. 21042v1 Announce Type: new Abstract: Autoregressive text-to-speech models achieve strong naturalness but suffer from slow inference due to sequential token generation, limiting their deployment in production applications that require low latency.
arXiv:2502.09245v3 Announce Type: replace Abstract: In contrast to RNNs, which compress their history into a single hidden state, Transformers can attend to all past tokens directly. However, standar...
Confucius4‑R2T2, a stable‑prefix streaming ASR model, is introduced to ensure that once a token is emitted it is never revised, thereby reducing wasted work and incorrect actions in real‑time transcription. The model is derived by post‑training Qwen3‑ASR‑1.7B on 1,500 hours of Chinese and English audio while keeping its audio encoder frozen, and it leverages a new concept called Longest Stable Prefix (LSP) to convert a full‑context model into a stable‑prefix streaming one without architectural changes. Evaluation shows that R2T2 achieves near full‑context quality with low latency and stable output, and its benefits generalize beyond the languages used in training.
arXiv:2605.26797v2 Announce Type: replace Abstract: We study Latent Recurrent Transformer (LRT), a lightweight augmentation of autoregressive transformers that reuses a high-level source-layer hidden...
arXiv:2608. 08569v1 Announce Type: new Abstract: Recent advancements in Speech Large Language Models have demonstrated remarkable capabilities in understanding complex audio tasks.
DeepSeek‑V4.1‑Flash is a multimodal Mixture‑of‑Experts model with 552 B backbone parameters that supports contexts of up to one million tokens. It uses a Causal Encoder‑Decoder architecture that activates 16 B parameters per token during decoding but only 8 B during prefill, improving cost efficiency for agentic workloads. The model introduces cross‑layer KV cache reuse via Compressed Sparse Attention 2 and FP4 KV caching, reducing its global KV cache footprint to 890 bytes per token and its persistent footprint to roughly one‑eighth of the previous version, while still delivering superior performance after pretraining on a 45 T‑token multimodal corpus.
The paper introduces acoustic-to-text KV compression for full‑duplex speech models, converting acoustic key‑value states into compact textual memory during listening‑time slack. When the KV cache exceeds a target budget, older acoustic states are evicted while transcripts and recent acoustic context are retained. Experiments on ten‑minute LongSpeech sessions show a 64.6% reduction in peak streaming KV‑cache size and improved transcription, temporal question answering, and summarization, with comparable pause‑handling, turn‑taking, and interruption performance in Full‑Duplex‑Bench.
arXiv:2607. 16339v1 Announce Type: new Abstract: Diffusion-based Large Language Models(DLLMs) enable parallel generation via Semi-Autoregressive (SAR) decoding in text generation.
BitSparse introduces a GPU‑resident, lossless compression format for sparse ReLU and ReLU² activations in feed‑forward networks. By integrating custom Triton kernels into an autograd function, it compresses and reconstructs activation caches without loss, supporting BF16 and FP8 formats and offering optional sign‑bit packing to approach theoretical limits. The method reduces per‑token VRAM usage by about 30% for Modded‑NanoGPT and Nemotron‑H, with negligible overhead and a new preallocated buffer system that eliminates host synchronization during training.
The paper introduces AWED, a word‑level emission‑delay metric, and demonstrates that post‑training a delayed‑streaming model with a joint reward (GRPO) improves both accuracy and latency. Using a single operating point (τ=6 frames), the method outperforms both its supervised baseline and the Voxtral Realtime backbone across all lookahead budgets, reducing WER by up to 30.8% at 80 ms delay and lowering median AWED from 1.17 s to 1.04 s.