W

Wav2vec2 Xlsr 1b Finnish Lm V2

Developed by aapot
A fine-tuned version of Facebook's wav2vec2-xls-r-1b model for Finnish automatic speech recognition tasks, trained on 275.6 hours of annotated Finnish speech data
Downloads 61
Release Time : 3/2/2022

Model Overview

This model is an automatic speech recognition (ASR) model for Finnish speech-to-text conversion, including an acoustic model and a KenLM language model, achieving a 4.09% word error rate on the Common Voice 7.0 test set

Model Features

High-performance Finnish recognition
Achieves a 4.09% word error rate and 0.88% character error rate on the Common Voice 7.0 test set
Large-scale pre-training foundation
Fine-tuned from the 1-billion-parameter wav2vec2-xls-r-1b model, which was pre-trained on 436,000 hours of multilingual data
Integrated language model
Includes a KenLM 5-gram language model specifically optimized for Finnish, significantly improving decoding performance
Diverse training data
Fine-tuned using 275.6 hours of Finnish data from various sources, including Common Voice, parliamentary sessions, broadcasts, and other scenarios

Model Capabilities

Finnish speech recognition
Short audio transcription (up to 20 seconds)
Speech decoding with language model

Use Cases

Speech-to-text
Meeting transcription
Automatically converts Finnish meeting recordings into text records
Suitable for formal speech with relatively high accuracy
Voice assistant
Provides speech recognition capabilities for Finnish voice assistants
Note the adaptability to informal speech
Media processing
Broadcast subtitle generation
Automatically generates subtitles for Finnish broadcast programs
Performs well with standard broadcast speech
Featured Recommended AI Models
AIbase
Empowering the Future, Your AI Solution Knowledge Base
Š 2025AIbase