SIFT-50M: A Large-Scale Multilingual Dataset for Speech Instruction Fine-Tuning

Fuente: arXiv
Salvato in:
Dettagli Bibliografici
Autori principali: Pandey, Prabhat, Swaminathan, Rupak Vignesh, Girish, K V Vijay, Sen, Arunasish, Xie, Jian, Strimel, Grant P., Schwarz, Andreas
Natura: Preprint
Pubblicazione: 2025
Soggetti:
Accesso online:
Tags: Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
_version_ 1866917988428414976
author Pandey, Prabhat
Swaminathan, Rupak Vignesh
Girish, K V Vijay
Sen, Arunasish
Xie, Jian
Strimel, Grant P.
Schwarz, Andreas
author_facet Pandey, Prabhat
Swaminathan, Rupak Vignesh
Girish, K V Vijay
Sen, Arunasish
Xie, Jian
Strimel, Grant P.
Schwarz, Andreas
contents We introduce SIFT (Speech Instruction Fine-Tuning), a 50M-example dataset designed for instruction fine-tuning and pre-training of speech-text large language models (LLMs). SIFT-50M is built from publicly available speech corpora, which collectively contain 14K hours of speech, and leverages LLMs along with off-the-shelf expert models. The dataset spans five languages, encompassing a diverse range of speech understanding as well as controllable speech generation instructions. Using SIFT-50M, we train SIFT-LLM, which outperforms existing speech-text LLMs on instruction-following benchmarks while achieving competitive performance on foundational speech tasks. To support further research, we also introduce EvalSIFT, a benchmark dataset specifically designed to evaluate the instruction-following capabilities of speech-text LLMs.
format Preprint
id arxiv_https___arxiv_org_abs_2504_09081
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle SIFT-50M: A Large-Scale Multilingual Dataset for Speech Instruction Fine-Tuning
Pandey, Prabhat
Swaminathan, Rupak Vignesh
Girish, K V Vijay
Sen, Arunasish
Xie, Jian
Strimel, Grant P.
Schwarz, Andreas
Audio and Speech Processing
Artificial Intelligence
Computation and Language
We introduce SIFT (Speech Instruction Fine-Tuning), a 50M-example dataset designed for instruction fine-tuning and pre-training of speech-text large language models (LLMs). SIFT-50M is built from publicly available speech corpora, which collectively contain 14K hours of speech, and leverages LLMs along with off-the-shelf expert models. The dataset spans five languages, encompassing a diverse range of speech understanding as well as controllable speech generation instructions. Using SIFT-50M, we train SIFT-LLM, which outperforms existing speech-text LLMs on instruction-following benchmarks while achieving competitive performance on foundational speech tasks. To support further research, we also introduce EvalSIFT, a benchmark dataset specifically designed to evaluate the instruction-following capabilities of speech-text LLMs.
title SIFT-50M: A Large-Scale Multilingual Dataset for Speech Instruction Fine-Tuning
topic Audio and Speech Processing
Artificial Intelligence
Computation and Language
url https://arxiv.org/abs/2504.09081