--- base_model: - Qwen/Qwen2.5-0.5B datasets: - openslr/librispeech_asr - slprl/SpokenSwag - slprl/sTinyStories library_name: transformers license: mit pipeline_tag: audio-to-audio --- # Slamming: Training a Speech Language Model on One GPU in a Day The model was presented in the paper [Slamming: Training a Speech Language Model on One GPU in a Day](https://arxiv.org/abs/2502.15814). # Paper abstract We introduce Slam, a recipe for training high-quality Speech Language Models (SLMs) on a single academic GPU in 24 hours. We do so through empirical analysis of model initialisation and architecture, synthetic training data, preference optimisation with synthetic data and tweaking all other components. We empirically demonstrate that this training recipe also scales well with more compute getting results on par with leading SLMs in a fraction of the compute cost. We hope these insights will make SLM training and research more accessible. In the context of SLM scaling laws, our results far outperform predicted compute optimal performance, giving an optimistic view to SLM feasibility. See code, data, models, samples at - https://pages.cs.huji.ac.il/adiyoss-lab/slamming . # Model Card for Model ID This is a Speech Language Model (SLM) trained for generating speech continuations over discrete [Hubert tokens](https://huggingface.co/slprl/mhubert-base-25hz). ## Model Details ### Model Description This Speech Language Model, introduced in ["_Slamming_: Training a Speech Language Model on One GPU in a Day"](https://arxiv.org/abs/2502.15814), focuses on efficient training. It was fine-tuned from [Qwen/Qwen2.5-0.5B](https://huggingface.co/Qwen/Qwen2.5-0.5B) over a vocabulary of 500 speech tokens extracted from the 11-th layer of [mhubert-25hz](https://huggingface.co/slprl/mhubert-base-25hz). The model was pre-trained using next-token prediction on a subset of LibriSpeech, Libri-Light and a synthetic dataset [sTinyStories](https://huggingface.co/datasets/slprl/sTinyStories). It was subsequently fine-tuned with DPO on [SpokenSwag](https://huggingface.co/datasets/slprl/SpokenSwag). - **Developed by:** [SLP-RL](https://huggingface.co/slprl) - **Model type:** SpeechLM - **License:** MIT - **Finetuned from model:** [Qwen/Qwen2.5-0.5B](https://huggingface.co/Qwen/Qwen2.5-0.5B) ### Model Sources - **Repository:** [https://github.com/slp-rl/slamkit](https://github.com/slp-rl/slamkit) - **Paper:** [https://arxiv.org/abs/2502.15814](https://arxiv.org/abs/2502.15814) - **Demo:** [https://pages.cs.huji.ac.il/adiyoss-lab/slamming/](https://pages.cs.huji.ac.il/adiyoss-lab/slamming/) ## Uses This base SpeechLM can be used to generate continuations for speech segments, or as a base for further tuning. See the _SlamKit_ [codebase](https://github.com/slp-rl/slamkit) for more details on usage, and checkout the [demo page](https://pages.cs.huji.ac.il/adiyoss-lab/slamming/) for some generation examples ### Out-of-Scope Use This model was trained on curated speech datasets which contain mainly audio-books and stories, as such the outputs should not be treated as factual in any way. ## How to Get Started with the Model We refer users to the official repository for full usage explanations - [github](https://github.com/slp-rl/slamkit). ## Training Details We highly encourage users to read the full [paper](https://arxiv.org/abs/2502.15814), for full training details, a brief overview is provided below. ### Training Data This model was trained on a subset of [LibriSpeech](https://huggingface.co/datasets/openslr/librispeech_asr) train, [Libri-Light](https://ai.meta.com/tools/libri-light/) and the synthetic dataset [sTinyStories](https://huggingface.co/datasets/slprl/sTinyStories) for the pre-training phase. It was also trained with DPO on the synthetic dataset [SpokenSwag](https://huggingface.co/datasets/slprl/SpokenSwag). ### Training Procedure This model was trained by next token prediction over several datasets, and then trained with DPO over [SpokenSwag](https://huggingface.co/datasets/slprl/SpokenSwag). Please refer to the [paper](https://arxiv.org/abs/2502.15814) or [code](https://github.com/slp-rl/slamkit) for the full training recipes. #### Preprocessing Speech tokens are extracted from the audio using [Hubert-25hz](https://huggingface.co/slprl/mhubert-base-25hz), and quantised using the official kmeans released with the model in [textlesslib](https://github.com/facebookresearch/textlesslib/tree/main). Units are de-duplicated. We encourage you to explore the official repository for full details - [github](https://github.com/slp-rl/slamkit). ## Evaluation The paper provides full results, we do give here some results and also refer to the [demo page](https://pages.cs.huji.ac.il/adiyoss-lab/slamming/) to listen to some samples. | Model | GPUs | Params | Num Tokens | sBLIMP ↑ | sStoryCloze ↑ | tStoryCloze ↑ | GenPPL ↓ | Auto-BLEU ↓ | |-------------------------------------------|---------|--------|---------------|-----------|---------------|---------------|----------|-------------| | **Speech only pre-training** | | | | | | | | | | GSLM | 8×V100 | 100M | 1B | 54.2 | 53.3 | 66.6 | — | — | | SyllableLM | 4×A40 | 300M | 16B | 63.7 | — | 75.4 | — | — | | TWIST-350M | 8×V100 | 305M | 10.8B | 56.2 | — | — | 137.3 | 3.46 | | TWIST-1.3B | 32×V100 | 1B | 10.8B | 57.0 | 52.4 | 70.6 | 131.8 | 3.20 | | TWIST-7B | 32×V100 | 7B | 36B | 59.0 | 55.3 | 74.1 | 93.74 | 3.06 | | TWIST-13B | 32×V100 | 13B | 36B | 59.2 | 55.4 | 76.4 | — | — | | Scaled Optimal | — | 823M | 82B | **61.3** | 56.7 | 78.0 | — | — | | Moshi | ?×H100 | 7B | ? | 58.9 | **58.7** | **81.8** | — | — | | SpiritLM | 64×A100 | 7B | 100B | 58.0 | 54.8 | 72.9 | — | — | | **With text / preference optimization** | | | | | | | | | | Scaling Interleaving | — | 9B | ~1T | — | **62.4** | 82.9 | — | — | | Moshi | ?×H100 | 7B | ~720B | 58.8 | 60.8 | 83.0 | — | — | | SpiritLM | 64×A100 | 7B | 100B | 58.3 | 61.0 | 82.9 | — | — | | AlignSLM-1.3B | 64×A100 | 1B | 10.8B + ~158B | 59.8 | 55.0 | 80.0 | — | — | | AlignSLM-7B | 64×A100 | 7B | 36B + ~158B | **62.3** | 61.1 | **86.8** | — | — | | **Ours (_Slam_)** | | | | | | | | | | _Slam_ (-DPO) | 2×A100 | 358M | 16.7B | 58.53 | 58.15 | 80.71 | 67.3 | 3.25 | | _Slam_ | 1×A5000 | 358M | 1.4B + 5M | 58.86 | 58.04 | 82.04 | 62.8 | 3.88 | | _Slam_ (scaled) | 2×A100 | 358M | 16.7B + 9M | **61.11** | **61.30** | **84.18** | **46.6** | 3.75 | ### Compute Infrastructure This model was trained as part of ["*Slamming*: Training a Speech Language Model on One GPU in a Day"](https://arxiv.org/abs/2502.15814), focusing on efficient training. #### Hardware This model was trained using **only 2 Nvidia A100 GPU** for **48 hours**. #### Software The model was trained using the [*SlamKit*](https://github.com/slp-rl/slamkit) codebase which builds upon 🤗transformers extending it to support easy and efficient training of Speech Language Models. ## Citation **BibTeX:** ``` @misc{maimon2025slamming, title={Slamming: Training a Speech Language Model on One GPU in a Day}, author={Gallil Maimon and Avishai Elmakies and Yossi Adi}, year={2025}, eprint={2502.15814}, archivePrefix={arXiv}, primaryClass={cs.LG}, url={https://arxiv.org/abs/2502.15814}, } ```