---
title: "TiCodec: Streaming Neural Speech Codecs through Time-Invariant Representations"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.05250v1"
permalink: "https://feed.astralor.com/items/itm_ad9ad5a7cc3bb7a6e92a6793"
published: "2026-07-06T15:57:49.000Z"
score: 77
channel: "tech"
tags: ["Dual-TIRE","TiCodec","TIRE","Streaming Speech Generation","Neural Speech Codec"]
---

# TiCodec: Streaming Neural Speech Codecs through Time-Invariant Representations

## AI commentary

> This paper not only improves codec efficiency but also reveals functional roles of different encoder layers through probing experiments, offering a theoretical basis for future efficient speech representations. Engineers working on voice assistants, real-time translation, and low-power edge devices will find TiCodec's architecture highly valuable.

## At a glance

TiCodec proposes a factorized representation that separates time-varying speech content from time-invariant information via the Time-Invariant Representation Extraction (TIRE) module, reducing frame-level modeling burden. Probing experiments reveal that intermediate encoder layers capture speaker/environment info with minimal linguistic content, and cross-file sampling improves robustness. Dual-TIRE, a multi-level architecture, leverages complementary encoder layers to enhance reconstruction quality and speaker similarity. Streaming inference with 660ms blocks shows negligible degradation, highlighting potential for low-latency speech generation systems.

## Key points

- Factorized representation separates time-varying speech from time-invariant info, reducing frame-level complexity
- Intermediate encoder layers capture speaker/environment features with nearly no linguistic content
- Cross-file sampling strategy significantly improves robustness of time-invariant representations
- Dual-TIRE multi-level architecture leverages complementarity of layers for better reconstruction and speaker similarity

## Full text

Neural speech codecs are increasingly used as intermediate representations in codec-based speech generation systems. TiCodec introduces a factorized representation that separates time-varying speech content from time-invariant information through a Time-Invariant Representation Extraction (TIRE) module, potentially reducing the amount of information that must be modeled at the frame-level. In this work, we investigate the nature of the information captured by TIRE representations and their suitability for low-latency speech processing. Using a series of probing tasks, we analyze the influence of the encoder layer and show that intermediate layers capture complementary speaker- and environment-related information while containing little linguistic content. We further study several segment selection strategies for TIRE training and demonstrate that cross-file sampling improves the robustness of invariant representations. Based on these findings, we propose Dual-TIRE, a multi-level architecture that exploits the complementarity of different encoder layers and improves speech reconstruction quality and speaker similarity. Finally, we evaluate TiCodec in a streaming inference setting using successive 660ms processing blocks. Results show that streaming operation can be achieved without significant degradation in reconstruction performance, highlighting the potential of factorized neural codec representations for future low-latency speech generation systems.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_ad9ad5a7cc3bb7a6e92a6793