The paper introduces a memory‑efficient audio synthesis architecture used in Siri Expressive Voices, converting semantic audio tokens into high‑fidelity audio via a residual vector quantization (RVQ) pipeline that decouples temporal and depth processing. A single Diffusion Transformer‑style depth decoder generates all RVQ levels autoregressively, while causal sliding‑window attention with fixed‑window key‑value caching keeps memory usage constant regardless of sequence length. Deployed on Apple’s AMX, the system achieves real‑time performance (≈10 ms per step) with only 21 MB peak runtime memory and 329 MB of on‑device assets, enabling continuous streaming of 20–320 s of audio and improving Mean Opinion Scores by +0.28 overall and +0.42 on conversational speech compared to the previous on‑device TTS system.
arXiv:2608. 08569v1 Announce Type: new Abstract: Recent advancements in Speech Large Language Models have demonstrated remarkable capabilities in understanding complex audio tasks.
The paper presents TTS‑Q, a post‑training quantization method that reduces the size of large diffusion transformer TTS models to 4‑bit precision while preserving speaker timbre and intelligibility. By identifying hard‑to‑quantize weights and activations and applying a mixed‑precision schedule that gives higher precision to late‑step feed‑forward activations, TTS‑Q restores timbre loss observed in standard 4‑bit PTQ. The approach also offers a calibrated‑W4A4 reference that lowers GPU memory usage, making TTS diffusion transformers more viable for on‑device, assistive, and real‑time applications.
The paper introduces Alignment-Free Text‑Audiobox (Text‑AB), a unified diffusion‑based framework that performs high‑quality voice dubbing and full‑duplex dialogue synthesis without requiring forced alignment. Text‑AB uses a latent diffusion model with DAC‑VAE features, achieving over 10× compression compared to prior EnCodec representations, and learns text‑speech alignment via cross‑attention. The authors pretrain a 3B‑parameter model on 480k hours of monolingual speech and fine‑tune it for cross‑lingual dubbing, full‑duplex dialogue, and emotional dialogue, reporting significant improvements in prosody, voice similarity, naturalness, and emotional expressivity over existing internal systems.
ZipCodec is a streaming neural speech codec that operates at an ultra‑low frame rate of 6.25 Hz and a bitrate of 0.80 kbps, achieving a theoretical latency of 160 ms. It leverages large‑scale WavLM distillation, a redesigned transformer architecture, a scalar spherical quantizer, and a latency‑aware streaming decoder to preserve reconstruction quality while reducing frame rate. Experiments demonstrate that ZipCodec outperforms existing streaming codecs at comparable bitrates in both reconstruction and downstream tasks, and it can run real‑time single‑stream inference on a consumer‑grade CPU despite having 842 M parameters.
The paper introduces TTS‑Q, a perception‑guided post‑training quantization method for zero‑shot text‑to‑speech diffusion transformers. It identifies that 4‑bit weight‑and‑activation quantization erodes speaker timbre, and addresses this by selecting hard‑to‑quantize items and applying a mixed‑precision schedule that gives higher precision to late‑step feed‑forward activations. The resulting approach restores timbre while preserving intelligibility and reduces GPU memory usage for resource‑constrained deployment.
The paper presents a redesign of the LiSenNet speech‑enhancement model for deployment on the STM32N6570‑DK Neural‑ART microcontroller accelerator. By replacing the recurrent bottleneck with convolutional mixers, converting unsupported operations to static int8 primitives, and using bounded decoder activations, the authors achieve an NPU‑compatible model that matches or surpasses the original LiSenNet in quality (PESQ 3.08 vs 3.01 FP32) while running each 16 ms input hop in 4.83 ms (real‑time factor 0.30). The study demonstrates that co‑designing parameter count, operator compatibility, quantization range, and streaming state is essential for efficient real‑time speech enhancement on constrained NPUs.
ZipCodec is a streaming neural speech codec that operates at an ultra‑low frame rate of 6.25 Hz and a bitrate of 0.80 kbps, achieving a theoretical latency of 160 ms. It leverages large‑scale WavLM distillation, a redesigned transformer architecture, a scalar spherical quantizer, and a latency‑aware streaming decoder. Experiments demonstrate that ZipCodec outperforms existing streaming codecs at comparable bitrates in both reconstruction quality and downstream tasks, while remaining real‑time on a consumer‑grade CPU despite its 842 M parameters.
The paper introduces a compression framework for the Whisper automatic speech recognition model that jointly optimizes six deployment dimensions—model size, temporal resolution, encoder token stride, low‑rank adaptation capacity, weight precision, and sparsity pattern—using NSGA‑III. The optimization targets three objectives: word error rate, inference FLOPs, and memory footprint. Evaluating 1,680 configurations, the study identifies compression combinations that outperform single‑axis scaling and notes that 1:4 structured sparsity cannot maintain acceptable accuracy within the tested budgets.
arXiv:2607. 29363v1 Announce Type: cross Abstract: Balancing sequence length, representational capacity, and long-horizon stability is a central problem in autoregressive (AR) speech and audio generation.
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.
FireRedAudio is a 9‑billion‑parameter audio language model that separates continuous input representations for audio understanding and speech generation, enabling a single autoregressive LLM to perform tasks such as ASR, zero‑shot TTS, Instruct TTS, and semantic/acoustic speech editing. The model uses a dedicated Audio Encoder for recognition and a RedAE‑based pathway for generation, with the LLM directly generating text or conditioning a flow‑matching DiT to produce acoustic latents. Evaluations show competitive or leading performance in multilingual ASR, content‑accurate zero‑shot TTS, strong instruction following, and significant improvements in speech editing over prior work.
StyleStream 2.0 is a fast, controllable streaming voice style conversion system that reduces end‑to‑end latency to about 520 ms on a consumer GPU, halving the delay of its predecessor while preserving conversion quality. It replaces mel‑spectrogram context with a compact style embedding, enabling a unified text‑conditioned flow model that supports both text‑based voice design and instruction‑based voice editing. Experiments demonstrate that StyleStream 2.0 achieves superior target style fidelity, voice design, and voice editing performance while maintaining intelligibility and attribute preservation.