AxForge
Sign in Get started

Model reference · open weights

wav2vec2-xls-r-italian

Audio dbdmg Speech→text 1 build Open weights 175k dl/mo

wav2vec2-xls-r-italian is an open-weight audio or speech model from dbdmg. wav2vec2-xls-r-300m-italian (FP32) weighs 631 MB; the smallest configuration that runs it is RTX 3060 12 GB.

wav2vec2-xls-r-italian is an automatic speech recognition model with 316M parameters, fine-tuned by dbdmg on the Common Voice Italian dataset. It is designed for processing Italian audio and is released under the Apache 2.0 license. The model achieves a word error rate of 0.1710 on its evaluation set.

Summary of the dbdmg/wav2vec2-xls-r-300m-italian model card, 2026-10-01

Hugging Face ↗

What it is

Released bydbdmg
TypeAudio & music
TaskSpeech→text
Parameters (lead)316M
Runs withtransformers
Released2022-03-02
Popularity175k downloads / month
Weights631 MB (wav2vec2-xls-r-300m-italian (FP32), file size)
LicenceOpen weights

What it runs on

Memory and cards for wav2vec2-xls-r-300m-italian (FP32)

Weights 631 MB (file size) · overhead about 1.6 GB.

CardOne streamCounted
memory
RTX 3060 12 GBfits11.6 GB
RTX 4060 Ti 16 GBfits15.4 GB
RTX 3090 24 GBfits23.4 GB
RTX 4090 24 GBfits23.4 GB
RTX 5090 32 GBfits31.0 GB
L40S 48 GBfits44.0 GB
A100 80 GBfits78.2 GB
H100 80 GBfits78.1 GB
RTX PRO 6000 Blackwell 96 GBfits93.8 GB
DGX Spark (GB10) 128 GB unifiedfits107 GB
H200 141 GBfits138 GB
B200 180 GBfits176 GB

Estimates, not measurements: the weights are the build's file size. A speech model's decoder keeps a small cache for every stream it transcribes, so memory grows with the streams and beams at once. Counted memory is 92 % of what CUDA reports for the card.

From the model card

What dbdmg says about wav2vec2-xls-r-italian

Read the model card

This model is a fine-tuned version of facebook/wav2vec2-xls-r-300m on the MOZILLA-FOUNDATION/COMMON_VOICE_7_0 - IT dataset. It achieves the following results on the evaluation set:

  • Loss: inf
  • Wer: 0.1710

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 0.0003
  • train_batch_size: 64
  • eval_batch_size: 8
  • seed: 42
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 500
  • num_epochs: 5.0
  • mixed_precision_training: Native AMP

Training results

Training LossEpochStepValidation LossWer
No log0.04100inf1.0
No log0.09200inf0.9983
No log0.13300inf0.7672
No log0.18400inf0.6919
2.99290.22500inf0.6266
2.99290.26600inf0.5513
2.99290.31700inf0.5081
2.99290.35800inf0.4945
2.99290.39900inf0.4720
0.53110.441000inf0.4387
0.53110.481100inf0.4411
0.53110.531200inf0.4429
0.53110.571300inf0.4322
0.53110.611400inf0.4532
0.46540.661500inf0.4492
0.46540.71600inf0.3879
0.46540.751700inf0.3836
0.46540.791800inf0.3743
0.46540.831900inf0.3687
0.42540.882000inf0.3793
0.42540.922100inf0.3766
0.42540.972200inf0.3705
0.42541.012300inf0.3272
0.42541.052400inf0.3185
0.39971.12500inf0.3244
0.39971.142600inf0.3082
0.39971.182700inf0.3040
0.39971.232800inf0.3028
0.39971.272900inf0.3112
0.36681.323000inf0.3110
0.36681.363100inf0.3067
0.36681.43200inf0.2961
0.36681.453300inf0.3081
0.36681.493400inf0.2936
0.36451.543500inf0.3037
0.36451.583600inf0.2974
0.36451.623700inf0.3010
0.36451.673800inf0.2985
0.36451.713900inf0.2976
0.36241.764000inf0.2928
0.36241.84100inf0.2860
0.36241.844200inf0.2922
0.36241.894300inf0.2866
0.36241.934400inf0.2776
0.35271.974500inf0.2792
0.35272.024600inf0.2858
0.35272.064700inf0.2767
0.35272.114800inf0.2824
0.35272.154900inf0.2799
0.31622.195000inf0.2673
0.31622.245100inf0.2962
0.31622.285200inf0.2736
0.31622.335300inf0.2652
0.31622.375400inf0.2551
0.30632.415500inf0.2680
0.30632.465600inf0.2558
0.30632.55700inf0.2598
0.30632.545800inf0.2518
0.30632.595900inf0.2541
0.29132.636000inf0.2507
0.29132.686100inf0.2500
0.29132.726200inf0.2435
0.29132.766300inf0.2376
0.29132.816400inf0.2348
0.27972.856500inf0.2512
0.27972.96600inf0.2382
0.27972.946700inf0.2523
0.27972.986800inf0.2522
0.27973.036900inf0.2409
0.27663.077000inf0.2453
0.27663.127100inf0.2326
0.27663.167200inf0.2286
0.27663.27300inf0.2342
0.27663.257400inf0.2305
0.24683.297500inf0.2238
0.24683.337600inf0.2321
0.24683.387700inf0.2305
0.24683.427800inf0.2174
0.24683.477900inf0.2201
0.24393.518000inf0.2133
0.24393.558100inf0.2217
0.24393.68200inf0.2189
0.24393.648300inf0.2105

|

Quoted from the model card on Hugging Face — the full card is behind the Hugging Face link above.

Benchmarks

Reported results

As published on the model card — the maker's own numbers, not measured by AxForge.

TaskDatasetMetricScore
Automatic Speech RecognitionCommon Voice 7Test WER19.440
Automatic Speech RecognitionCommon Voice 7Test CER4.470
Automatic Speech RecognitionCommon Voice 7Test WER (+LM)14.080
Automatic Speech RecognitionCommon Voice 7Test CER (+LM)3.670
Automatic Speech RecognitionRobust Speech Event - Dev DataTest WER31.010
Automatic Speech RecognitionRobust Speech Event - Dev DataTest CER9.270
Automatic Speech RecognitionRobust Speech Event - Dev DataTest WER (+LM)22.090
Automatic Speech RecognitionRobust Speech Event - Dev DataTest CER (+LM)7.900
Automatic Speech RecognitionRobust Speech Event - Test DataTest WER38.070
© 2026 AxForge · EU-hosted AI infrastructure Pricing Docs Trust Privacy Terms