| --- |
| license: apache-2.0 |
| language: |
| - pt |
| base_model: |
| - openai/whisper-large-v3 |
| library_name: transformers |
| tags: |
| - pt_PT |
| --- |
| |
| # Neo_EP |
| |
| Fine-tuned version of [`inesc-id/WhisperLv3-FT`](https://huggingface.co/inesc-id/WhisperLv3-FT) for **European Portuguese** automatic speech recognition, developed by [Neoscopio](https://huggingface.co/Neoscopio-SA). |
| |
| - **Architecture:** Transformer encoder-decoder (1550M parameters) |
| - **Base model:** [`inesc-id/WhisperLv3-FT`](https://huggingface.co/inesc-id/WhisperLv3-FT) (from [`openai/whisper-large-v3`](https://huggingface.co/openai/whisper-large-v3)) |
| - **Language:** European Portuguese (`pt`) |
| - **Task:** Transcription |
| - **Compute type:** float16 |
| |
| > **Note:** A paper with full training methodology, evaluation results, and benchmarks is currently under preparation and will be published soon. |
| |
| **Current results:** |
| |
| |
| | # | Modelo | WER (%) | CER (%) | RTF | Tempo (s) | eurospeech | falabracarense | MLS | |
| |---| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: | |
| | 1 | Neo_EP | **13.67%** | 10.16% | 0.000 | 1208.3s | **28.1%** | **7.1%** | 5.8% | |
| | 2 | EP-X(Faster-Whisper) | 18.93% | 14.56% | 0.000 | 1259.1s | 42.5% | 7.6% | 6.6% | |
| | 3 | whisper-large-v3 | 26.99% | 19.57% | 0.000 | 1202.9s | 41.8% | 33.9% | **5.2%** | |
| | 4 | Nvidia-Canary-1b-v2 | 32.07% | 22.06% | 0.000 | 2296.4s | 45.6% | 43.5% | 7.1% | |
