--- license: apache-2.0 tags: - whisper-event - generated_from_trainer datasets: - google/fleurs metrics: - wer model-index: - name: Whisper Small Amharic FLEURS results: - task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: google/fleurs am_et type: google/fleurs config: am_et split: test+validation args: am_et metrics: - name: Wer type: wer value: 160.99 language: - amh --- # Whisper Small Tamil FLEURS This model is a fine-tuned version of [openai/whisper-small](https://huggingface.co/openai/whisper-small) on the google/fleurs am_et dataset. It achieves the following results on the evaluation set: - Loss: - - Wer: 160.99 ## Model description - The main Whisper Small Hugging Face page: [Hugging Face - Whisper Small](https://huggingface.co/openai/whisper-small) ## Intended uses & limitations - For experimentation and curiosity. - Based on the paper [AXRIV](https://arxiv.org/abs/2212.04356) and [Benchmarking OpenAI Whisper for non-English ASR - Dan Shafer](https://blog.deepgram.com/benchmarking-openai-whisper-for-non-english-asr/), there is a performance bias towards certain languages and curated datasets. - From the Whisper paper, am_et is a low resource language (Table E), with the WER results ranging from 120-229, based on model size. Whisper small WER=120.2, indicating more training time may improve the fine tuning. ## Training and evaluation data - This model was trained/evaluated on "test+validation" data from google/fleurs [google/fluers - HuggingFace Datasets](https://huggingface.co/datasets/google/fleurs). ## Training procedure - The training was done in Lambda Cloud GPU on A100/40GB GPUs, which were provided by OpenAI Community Events [Whisper Fine Tuning Event - Dec 2022](https://github.com/huggingface/community-events/tree/main/whisper-fine-tuning-event#fine-tune-whisper). The training was done using [HuggingFace Community Events - Whisper - run_speech_recognition_seq2seq_streaming.py](https://github.com/huggingface/community-events/blob/main/whisper-fine-tuning-event/run_speech_recognition_seq2seq_streaming.py) using the included [whisper_python_am_et.ipynb](https://huggingface.co/drmeeseeks/whisper-small-am_et/blob/main/am_et_fine_tune_whisper_streaming_colab_RUNNING-evalerrir.ipynb) to setup the Lambda Cloud GPU/Colab environment. For Colab, you must reduce the train batch size to the recommended amount mentioned at , as the T4 GPUs have 16GB of memory [Whisper Fine Tuning Event - Dec 2022](https://github.com/huggingface/community-events/tree/main/whisper-fine-tuning-event#fine-tune-whisper). The notebook sets up the environment, logs into your huggingface account, and generates a bash script. The bash script generated in the IPYNB, `run.sh` was run from the terminal to train `bash run.sh`, as described on the Whisper community events GITHUB page. Num Examples = 446, Num Epochs = 715, Number of trainable parameters = 241734912. ### Training hyperparameters The following hyperparameters were used during training: - learning_rate: 1e-05 - train_batch_size: 64 - eval_batch_size: 32 - seed: 42 - gradient_accumulation_steps: 1 - total_train_batch_size: 64 - optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08 - lr_scheduler_type: linear - lr_scheduler_warmup_steps: 500 - training_steps: 1000 - mixed_precision_training: Native AMP - do_eval=False ### Training results 5000 Steps | Training Loss | Epoch | Step | Learning Rate | |:-------------:|:-----:|:-----:|:-----:| | 3.0968 | 0.05 | - | 4.2e-7 | | 1.178 | 28.57 | - | 3.92e-6 | | 0.03 | 53.57 | - | 7.42e-6 | | 0.0002 | 217.86 | - | 7.73e-6 | | 0.0001 | 378.57 | ~ 2000 | 5.23e-6 | | 0.0000 | 382.14 | - | 5.14e-6 | | 0.0000 | 467.86 | 3300 | 3.84e-6| | 0.0000 | 614.29 | 4300 | 1.56e-6| | 0.0000 | 685.71 | 4812 | 4.53e-7| | 0.0000 | 710.71 | 4997 | 6.44e-8| 3000 Steps | Training Loss | Epoch | Step | Learning Rate | |:-------------:|:-----:|:-----:|:-----:| | 3.0968 | 0.05 | - | 4.2e-7 | | 0.0017 | 96.43 | 687 | 9.316e-6 | ### Recommendations Limit training duration for smaller datasets to ~ 2000 to 3000 steps to avoid overfitting. 5000 steps using the [HuggingFace - Whisper Small](https://huggingface.co/openai/whisper-small) takes ~ 5hrs on A100 GPUs (1hr/1000 steps). Encountered `RuntimeError: The size of tensor a (504) must match the size of tensor b (448) at non-singleton dimension 1` which is related to [Trainer RuntimeError](https://discuss.huggingface.co/t/trainer-runtimeerror-the-size-of-tensor-a-462-must-match-the-size-of-tensor-b-448-at-non-singleton-dimension-1/26010) as some languages datasets have input lengths that have non-standard lengths. The link did not resolve my issue, and appears elsewhere too [Training languagemodel – RuntimeError the expanded size of the tensor (100) must match the existing size (64) at non singleton dimension 1](https://hungsblog.de/en/technology/troubleshooting/training-languagemodel-runtimeerror-the-expanded-size-of-the-tensor-100-must-match-the-existing-size-64-at-non-singleton-dimension-1/). To circumvent this issue, `run.sh` only trains and saves the model (if you make changes to `run.sh` be sure to clear/rm the contents as piping appends). Then run `python run_eval_whisper_streaming.py --model_id="openai/whisper-small" --dataset="google/fleurs" --config="am_et" --batch_size=32 --max_eval_samples=64 --device=0 --language="am"` to find the WER score. Erroring out during evaluation prevents the trained model from loading to HugginFace. Based on the paper [AXRIV](https://arxiv.org/abs/2212.04356) and [Benchmarking OpenAI Whisper for non-English ASR - Dan Shafer](https://blog.deepgram.com/benchmarking-openai-whisper-for-non-english-asr/), there is a performance bias towards certain languages and curated datasets. The OpenAI fintuning community event provided ample _free_ GPU time to help develop the model further and improve WER scores. ### Environmental Impact Carbon emissions were estimated using the [Machine Learning Impact calculator](https://mlco2.github.io/impact#compute) presented in [Lacoste et al. (2019)](https://arxiv.org/abs/1910.09700). In total roughly 100 hours were used primarily in US East/Asia Pacific (80%/20%), with AWS as the reference. Additional resources are available at [Our World in Data - CO2 Emissions](https://ourworldindata.org/co2-emissions) - __Hardware Type__: AMD EPYC 7J13 64-Core Processor (30 core VM) 197GB RAM, with NVIDIA A100-SXM 40GB - __Hours Used__: 100 hrs - __Cloud Provider__: Lambda Cloud GPU - __Compute Region__: US East/Asia Pacific - __Carbon Emitted__: 12 kg (GPU) + 13 kg (CPU) = 25 kg (the weight of 3 gallons of water) ### Citation - [Whisper - GITHUB](https://github.com/openai/whisper) - [Whisper - OpenAI - BLOG](https://openai.com/blog/whisper/) - [Model Card - HuggingFace Hub - GITHUB](https://github.com/huggingface/huggingface_hub/blob/main/src/huggingface_hub/templates/modelcard_template.md) ```bibtex @misc{https://doi.org/10.48550/arxiv.2212.04356, doi = {10.48550/ARXIV.2212.04356}, url = {https://arxiv.org/abs/2212.04356}, author = {Radford, Alec and Kim, Jong Wook and Xu, Tao and Brockman, Greg and McLeavey, Christine and Sutskever, Ilya}, keywords = {Audio and Speech Processing (eess.AS), Computation and Language (cs.CL), Machine Learning (cs.LG), Sound (cs.SD), FOS: Electrical engineering, electronic engineering, information engineering, FOS: Electrical engineering, electronic engineering, information engineering, FOS: Computer and information sciences, FOS: Computer and information sciences}, title = {Robust Speech Recognition via Large-Scale Weak Supervision}, publisher = {arXiv}, year = {2022}, copyright = {arXiv.org perpetual, non-exclusive license} } @article{owidco2andothergreenhousegasemissions, author = {Hannah Ritchie and Max Roser and Pablo Rosado}, title = {CO₂ and Greenhouse Gas Emissions}, journal = {Our World in Data}, year = {2020}, note = {https://ourworldindata.org/co2-and-other-greenhouse-gas-emissions} } ``` ### Framework versions - Transformers 4.26.0.dev0 - Pytorch 1.13.0+cu117 - Datasets 2.7.1.dev0 - Tokenizers 0.13.2