--- language: - ur library_name: nemo datasets: - mozilla-foundation/common_voice_12_0 thumbnail: null tags: - automatic-speech-recognition - speech - audio - Transducer - FastConformer - Conformer - pytorch - NeMo license: cc-by-4.0 widget: - Title: Common Voice Urdu Sample src: https://cdn-media.huggingface.co/speech_samples/sample_urdu.flac model-index: - name: parakeet-rnnt-0.6b-urdu results: - task: name: Automatic Speech Recognition type: automatic-speech-recognition dataset: name: Mozilla Common Voice 12.0 (Urdu) type: mozilla-foundation/common_voice_12_0 split: test args: language: ur metrics: - name: Test WER type: wer value: 25.513 metrics: - wer pipeline_tag: automatic-speech-recognition --- # Fine-Tuned Parakeet RNNT 0.6B (Urdu) This repository contains the fine-tuned version of the **Parakeet RNNT 0.6B** model for **Urdu** Automatic Speech Recognition (ASR). The base model, developed by **NVIDIA NeMo** and **Suno.ai**, was fine-tuned on the Urdu dataset from Mozilla's Common Voice 12.0. This fine-tuning enables the model to perform speech-to-text tasks in Urdu with improved accuracy and domain-specific adaptation. --- ## Model Overview The **Parakeet RNNT** is an XL version of the FastConformer Transducer with **600 million parameters**, optimized for ASR tasks. The fine-tuned model supports Urdu transcription, enabling applications such as subtitling, speech analytics, and voice-assisted interfaces. Base model details can be found on 🤗 [Hugging Face](https://huggingface.co/nvidia/parakeet-rnnt-0.6b). --- ## Training Details ### Dataset The fine-tuning was performed using the **Urdu dataset** from Mozilla's [Common Voice 12.0](https://huggingface.co/datasets/mozilla-foundation/common_voice_12_0). This dataset provides diverse speech samples in Urdu, ensuring robust training. ### Hardware - **Google Colab Pro** - **NVIDIA A100 GPU** --- ## Results The model achieved a **Word Error Rate (WER)** of **25.513%** on the test split of the Common Voice Urdu dataset. While this may seem high, the model demonstrates impressive accuracy in many transcriptions: - **Reference**: کچھ بھی ہو سکتا ہے۔ **Predicted**: کچھ بھی ہو سکتا ہے۔ --- - **Reference**: اورکوئی جمہوریت کو کوس رہا ہے۔ **Predicted**: اور کوئ جمہوریت کو کو س رہا ہے۔ This WER is slightly higher than OpenAI's **Whisper model**, which achieved **23%** without fine-tuning ([reference](https://arxiv.org/html/2409.11252v1)), but demonstrates the potential of the Parakeet RNNT with further fine-tuning. --- ## How to Use this Model ### Loading the Model You can load the fine-tuned model using NVIDIA NeMo: ```python import nemo.collections.asr as nemo_asr asr_model = nemo_asr.models.EncDecRNNTBPEModel.from_pretrained(model_name="hash2004/parakeet-fine-tuned-urdu") ``` ## How to Fine Tune this Model You can find all resources on fine-tuning the Parakeet RNNT (0.6B) model on [this GitHub Repository](https://github.com/hash2004/conformer-fine-tuned-urdu).