Entity · paper

Content is What Remains: Invariant Speech Tokenization from Parallel Utterances

paperactiveprovisionalcontent-is-what-remains-invariant-speech-tokenization-from-parallel-utterances-8cdb4ee5·1 events·first seen Jul 22, 2026

Aliases: Content is What Remains: Invariant Speech Tokenization from Parallel Utterances

Co-occurring entities

HuBERT PINT (Parallel INvariant Tokenization)

More like this (12)

Interleaved Speech Language Models Latently Work In Text How Surprising Is Historical Italian to Language Models? Tokenization Tax, Comprehension Tax, and a Simple Mitigation PINT (Parallel INvariant Tokenization)In-Place Tokenizer Expansion for Pre-trained LLMs Partition, Prompt, Aggregate: Statistical Self-Consistency in Language Models First-Token Broadcasters: Mechanistic Origins of Language Identity and Distributed Robustness in Transformers Beyond Uniform Tokens: Adaptive Compression for Time Series Language Models Rate-Utility Frontiers for Language Encodings: Comparing Tokens, Bytes, and Pixels Under Controlled Linguistic Content intra-frame token sparsification Beyond Tokenization: Direct Timestep Embedding and Contrastive Alignment for Time-Series Question Answering Unigram tokenisation Tokenizers

Recent events (1)

5arXiv · cs.CL·Jul 22, 2026·source ↗

PINT: Parallel utterance training yields near-speaker-invariant discrete speech tokens

Researchers introduce PINT (Parallel INvariant Tokenization), a method that fine-tunes a self-supervised speech encoder using alignment losses across parallel utterances from multiple speakers to extract purely linguistic content tokens. The approach achieves a 98.7% relative reduction in speaker probe accuracy and 42% lower ABX error rate compared to baselines, while preserving frame-level temporal grounding. PINT tokens are designed as drop-in semantic targets for audio codecs, addressing a known limitation of HuBERT-style SSL models that leak speaker identity and prosody into discrete representations.

Evaluation and Benchmarking Multimodal Progress Content is What Remains: Invariant Speech Tokenization from Parallel Utterances HuBERT PINT (Parallel INvariant Tokenization)