# Common-voice + librispeech

**URL:** </static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fcommon-voice-librispeech%2F19782%3E%5C
**Category:** Beginners\
**Created:** [June 29, 2022, 1:38pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fcommon-voice-librispeech%2F19782 "2022-06-29T13:38:08Z")\
**Posts on this page:** 1\
**Page:** 1

<div class="post-metadata">

**Author:** ![gciaffoni](https://avatars.discourse-cdn.com/v4/letter/g/8baadc/32.png) [@gciaffoni](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Fu%2Fgciaffoni)%5C
**Post date:** [June 29, 2022, 1:38pm UTC](/static-proxy?url=https%3A%2F%2Fdiscuss.huggingface.co%2Ft%2Fcommon-voice-librispeech%2F19782%2F1 "2022-06-29T13:38:08Z")

</div>

I’ve recently been experimenting with deep learning with wav2vec2 for ASR, I made my first fine-tuning with common-voice in Italian reaching a wer equal to 12%, now I thought I’d take the last checkpoint and continue fine-tuning with librispeech in Italian to further lower the wer. It’s possible?  
I take the same notebook I load the librispeech dataset I get the same vocabulary and then I load the model from the last checkpoint: “model = Wav2Vec2ForCTC.from\_pretrained (‘model\_checkpoint,’ …)”.  
I ask this question because I notice that the WER starts at 22% at the first checkpoints … maybe it’s not a correct approach, maybe it was better to stay on common-voice and increase the epoch.

thank you all for your attention.
