|
--- |
|
language: |
|
- la |
|
license: agpl-3.0 |
|
tags: |
|
- robust-speech-event |
|
- hf-asr-leaderboard |
|
datasets: |
|
- lsb/poetaexmachina-mp3-recitations |
|
metrics: |
|
- wer |
|
model-index: |
|
- name: wav2vec2-base-it-latin |
|
results: |
|
- task: |
|
type: automatic-speech-recognition |
|
name: Speech Recognition |
|
dataset: |
|
type: lsb/poetaexmachina-mp3-recitations |
|
name: Poeta Ex Machina mp3 recitations |
|
metrics: |
|
- type: wer |
|
value: 0.398 |
|
name: Test WER |
|
--- |
|
--- |
|
|
|
# wav2vec2-base-it-latin |
|
|
|
This model is a fine-tuned version of [wav2vec2-base-it-voxpopuli](https://huggingface.co/facebook/wav2vec2-base-it-voxpopuli) |
|
|
|
The dataset used is the [poetaexmachina-mp3-recitations](https://github.com/lsb/poetaexmachina-mp3-recitations), |
|
all of the 2-series texts (vergil) and every tenth 1-series text (words from Poeta Ex Machina's [database](https://github.com/lsb/poetaexmachina/blob/master/merged-scansions.db) of words with scansions). |
|
|
|
It achieves the following [results](https://github.com/lsb/tironiculum/blame/trunk/wav2vec2%20base%20it%20latin.ipynb#L1234) on the evaluation set: |
|
|
|
- Loss: 0.1943 |
|
- WER: 0.398 |
|
|
