Meta Researchers Present GSLM2 Speech Model

Meta researchers presented GSLM2 on November 04, 2026, to improve spoken language models that learn linguistic structure from speech alone. They say these models have lagged text-based language models in capability and data-scaling efficiency. The team identifies two causes: existing speech units have limited linguistic quality, and systems rely on a single stream of discrete units. Multiple codebooks are commonly used to encode acoustic features, but rarely to represent speech meaning.

GSLM2 learns complementary speech properties at multiple abstraction levels. Its SpidR 2.0 encoder uses self-supervised learning and differentiable prediction heads to produce discrete units at those levels. A multi-stream spoken language model jointly models the units with a shared temporal backbone, predictors dedicated to each stream, and an autoregressive head that captures dependencies between streams.

Across training sets of up to 6M hours and models with up to 6.5B parameters, GSLM2 consistently improved linguistic competence and scaled more efficiently with data, Meta says. On read and conversational speech, it narrowed the gap with text-based language models by about two orders of magnitude on sWuggy, sBLIMP, and tSC, and outperformed competing speech-text systems on all three tasks. The researchers released the training code and model checkpoints as open source on GitHub; the work was published on arXiv.

The paper names Yosuke Higuchi, Tom Fizycki, Mahi Luthra, Charles-Éric Saint-James, Azeez Syed, Angel Villar Corrales, Mido Assran, Nicolas Ballas, Jiayi Shen, Dongyan Lin, Phillip Rust, Surya Parimi, Vanessa Stark, Alexandre Allauzen, Balázs Kegl, Juan Pino, and Emmanuel Dupoux as authors. Meta’s research page lists Speech & Audio as the topic.

Meta researchers presented GSLM2 on November 04, 2026, to improve spoken language models that learn linguistic structure from speech alone. They say these models have lagged text-based language models in capability and data-scaling efficiency. The team identifies two causes: existing speech units have limited linguistic quality, and systems rely on a single stream of discrete units. Multiple codebooks are commonly used to encode acoustic features, but rarely to represent speech meaning.

GSLM2 learns complementary speech properties at multiple abstraction levels. Its SpidR 2.0 encoder uses self-supervised learning and differentiable prediction heads to produce discrete units at those levels. A multi-stream spoken language model jointly models the units with a shared temporal backbone, predictors dedicated to each stream, and an autoregressive head that captures dependencies between streams.

Across training sets of up to 6M hours and models with up to 6.5B parameters, GSLM2 consistently improved linguistic competence and scaled more efficiently with data, Meta says. On read and conversational speech, it narrowed the gap with text-based language models by about two orders of magnitude on sWuggy, sBLIMP, and tSC, and outperformed competing speech-text systems on all three tasks. The researchers released the training code and model checkpoints as open source on GitHub; the work was published on arXiv.

The paper names Yosuke Higuchi, Tom Fizycki, Mahi Luthra, Charles-Éric Saint-James, Azeez Syed, Angel Villar Corrales, Mido Assran, Nicolas Ballas, Jiayi Shen, Dongyan Lin, Phillip Rust, Surya Parimi, Vanessa Stark, Alexandre Allauzen, Balázs Kegl, Juan Pino, and Emmanuel Dupoux as authors. Meta’s research page lists Speech & Audio as the topic.