Whisper Large V3 Turbo

🚀 Model Card for Model ID

This model is a Hebrew finetune (continued training) of the OpenAI Whisper Large v3 Turbo model, designed for Hebrew audio transcription.

🚀 Quick Start

Please follow the original model card for usage details - replacing with this model name. You can also find other weight formats and quantizations on the ivrit ai HF page.

We created some simple example scripts using this model and weights for other inference runtimes. Find those in the "examples" folder within the training GitHub repo.

✨ Features

• This model is a Hebrew finetune of the OpenAI Whisper Large v3 Turbo model.
• It is intended for mostly - Hebrew audio transcription.

📚 Documentation

Model Details

Model Description

Bias, Risks, and Limitations

⚠️ Important Note

The language detection capability of this model has been degraded during training - it is intended for mostly - Hebrew audio transcription. The language token should be explicitly set to Hebrew.

Additionally, the translation task was not trained and also degraded. This model would not be able to translate in any reasonable capacity.

Training Details

Training Data

This model was trained on the following datasets:

• ivrit-ai/crowd-transcribe-v5 - Publicly accessible audio sources have been crowd - transcribed segment - by - segment (~300h).
• ivrit-ai/crowd-recital-whisper-training - Crowd - sourced recording of Wikipedia article snippets (~50h).
• ivrit-ai/knesset-plenums-whisper-training - A subset of a Knesset (Israeli house of representatives) plenum protocols (~325h).

Training Procedure

This model is a weighted - average of the lowest eval loss checkpoints (From around the end of epoch 2) from two separate runs with the same setup. Training code can be found on the ivrit - ai Github here

Preprocessing

The "Crowd Recital" and "Knesset" datasets contain timestamps and previous text following the Whisper expected inputs. Timestamps were used from 40% of samples from those datasets, and 50% of the previous text was used.

The "Crowd Transcribe" datasets has no timestamps or previous text and this preprocessing only included melspec feature extraction and text encoding.

Preprocessing code can be found within the training code repository.

Datasets were interleaved with 0.15:0.8:0.05 ratio (knesset:crowd - transcribe:crowd - recital).

Training Hyperparameters

• Training regime: bf16 mixed precision with sdpa
• Learning Rate: 1e - 5, Linear decay, 800 steps warmup for 3 epochs
• Batch Size: 32

Training Hardware / Duration

• GPU Type: 8 x Nvidia A40 machine
• Duration: ~9h run, stopped at 3 epochs

Evaluation

Please refer to the ivrit-ai/hebrew-transcription-leaderboard

📄 License

This model is licensed under the Apache - 2.0 license.