|
--- |
|
license: mit |
|
datasets: |
|
- rsalshalan/QASR |
|
- DynamicSuperb/DialectIdentification_ADI17 |
|
language: |
|
- ar |
|
- en |
|
metrics: |
|
- bleu |
|
- wer |
|
- accuracy |
|
base_model: |
|
- deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B |
|
pipeline_tag: audio-text-to-text |
|
--- |
|
# TinyOctopus: Bilingual Audio Language Model 🐙🔊 |
|
|
|
## 📢 Overview |
|
**TinyOctopus** is a **Bilingual Audio Language Model (Audio-LLM)** designed to process and generate text from audio inputs. The model leverages **Distil-Whisper (distil-large-v3)** for audio encoding, a **cross-attention projection layer** for alignment, and **DeepSeek 1.5B** for text generation. TinyOctopus is optimized for tasks such as: |
|
|
|
- **Bilingual Automatic Speech Recognition (ASR)** 🗣️ |
|
- **Speech Translation** 🌍 |
|
- **Dialect Identification** |
|
|
|
TinyOctopus maintaining the architectural principles of the following structure: |
|
|
|
## 🏗 Model Architecture |
|
### **TinyOctopus integrates:** |
|
1. **Distil-Whisper (distil-large-v3)** for encoding audio inputs. |
|
2. **Cross-Attention Projection Layer** (trainable) to align audio features with textual representations. |
|
3. **DeepSeek 1.5B** as the core language model for text generation. |
|
|
|
## 📂 Dataset |
|
The model has been trained on multiple datasets to optimize its performance across different tasks: |
|
|
|
- **[QASR Dataset](https://arxiv.org/pdf/2106.13000)**: QASR is the largest transcribed Arabic speech corpus, collected from the broadcast domain. It contains **2,000 hours of multi-dialect speech** sampled at **16kHz** from **Al Jazeera News Channel**, with lightly supervised transcriptions aligned with the audio segments. Unlike previous datasets, QASR includes **linguistically motivated segmentation, punctuation, speaker information**, and more. The dataset is suitable for **ASR, Arabic dialect identification, punctuation restoration, speaker identification, and NLP applications**. Additionally, a **130M-word language model dataset** is available to aid language modeling. Speech recognition models trained on QASR achieve competitive **WER** compared to the MGB-2 corpus, and it has been used for downstream tasks like **Named Entity Recognition (NER)** and **punctuation restoration**. |
|
|
|
- **[ADI17 Dataset](https://swshon.github.io/pdf/shon_2020_adi17.pdf)**: ADI17 is a **large-scale Arabic Dialect Identification (DID) dataset**, collected from **YouTube videos** across **17 Arabic-speaking countries in the Middle East and North Africa**. It contains **3,000 hours of speech** for training DID systems and an additional **57 hours** for development and testing. The dataset is categorized into **short (<5s), medium (5-20s), and long (>20s) speech segments** for detailed evaluation. ADI17 enables state-of-the-art **dialect identification** and provides a robust evaluation platform. It has been benchmarked on **domain-mismatched conditions** using the Multi-Genre Broadcast 3 (MGB-3) test set. |
|
|
|
## ⚙️ Installation & Usage |
|
### **💻 Install Dependencies** |
|
```bash |
|
pip install -r requirements.txt |
|
|
|
|
