GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
crslnjoyz
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
crslnjoyz
Model Tree
openai/whisper-large-v3
Input Modalities
Output Modalities
Supported Functionality
This model is a fine-tuned version of openai/whisper-large-v3, adapted using Parameter-Efficient Fine-Tuning (PEFT) with Low-Rank Adaptation (LoRA) for Automatic Speech Recognition (ASR) of the low-resource Tboli language.
It is designed to transcribe spoken Tboli words and short phrases into text using a predefined vocabulary. The model serves as the speech recognition component of a Tboli-to-English speech translation system.
This repository contains LoRA adapter weights intended to be used with the original Whisper Large V3 base model.
The model was fine-tuned on a privately maintained Tboli speech dataset containing audio recordings of predefined words and short phrases paired with their corresponding transcriptions. Audio recordings were preprocessed and resampled to 16 kHz. Data augmentation techniques, including noise, pitch, and speed variations, were applied to the training samples. The dataset was divided into training, validation, and testing subsets using an 80/10/10 split. Evaluation was conducted using Character Error Rate (CER) and Word Error Rate (WER).
Dataset availability: The datset is private and not publicly available. The release of this model does not grant access to, ownership of, or permission to redistribute the underlying Tboli speech dataset.
These results reflect speech recognition performance within the predefined Tboli vocabulary used during training and evaluation.
Fine-tuning method
The following hyperparameters were used during training:
Given the limited vocabulary coverage and the recording-level dataset split, additional evaluations would be necessary to establish model generalization beyond the predefined research corpus.
Future evaluation should consider unseen vocabulary, continuous speech, additional native speakers, and more diverse acoustic conditions.
The dataset and its evaluation results should be understood as an initial contribution toward low-resource Tboli speech recognition research rather than a comprehensive representation of the Tboli language.