---
library_name: transformers
license: mit
language: it
metrics:
- per
tags:
- audio
- automatic-speech-recognition
- speech
- phonemize
- phoneme
datasets:
- facebook/multilingual_librispeech
model-index:
- name: Wav2Vec2-base Italian finetuned for phonemes by LMSSC
  results:
  - task:
      type: automatic-speech-recognition
      name: Speech Recognition
    dataset:
      name: Multilingual Librispeech
      type: facebook/multilingual_librispeech
      args: it
    metrics:
    - type: per
      value: 4.34
      name: Test PER on Multilingual Librispeech IT | Trained
    - type: per
      value: 4.25
      name: Val PER on Multilingual Librispeech IT | Trained
---

# Fine-tuned Italian Voxpopuli v2 wav2vec2-base model for speech-to-phoneme task in Italian

Fine-tuned [facebook/wav2vec2-base-it-voxpopuli-v2](https://huggingface.co/facebook/wav2vec2-base-it-voxpopuli-v2) for **Italian speech-to-phoneme** (without language model) using the train and validation splits of [Multilingual Librispeech](https://huggingface.co/datasets/facebook/multilingual_librispeech).

## Audio samplerate for usage 

When using this model, make sure that your speech input is **sampled at 16kHz**.

## Output

As this model is specifically trained for a speech-to-phoneme task, the output is sequence of [IPA-encoded](https://en.wikipedia.org/wiki/International_Phonetic_Alphabet) words, without punctuation.
If you don't read the phonetic alphabet fluently, you can use this excellent [IPA reader website](http://ipa-reader.xyz) to convert the transcript back to audio synthetic speech in order to check the quality of the phonetic transcription.

## Training procedure

The model has been finetuned on Multilingual Librispeech (IT) for 30 epochs on a 1xADA_6000 GPU at Cnam/LMSSC using a ddp strategy and gradient-accumulation procedure (256 audios per update, corresponding roughly to 25 minutes of speech per update -> 2k updates per epoch)

- Learning rate schedule : Double Tri-state schedule
    - Warmup from 1e-5 for 7% of total updates
    - Constant at 1e-4 for 28% of total updates
    - Linear decrease to 1e-6 for 36% of total updates
    - Second warmup boost to 3e-5 for 3% of total updates
    - Constant at 3e-5 for 12% of total updates
    - Linear decrease to 1e-7 for remaining 14% of updates
 
- The set of hyperparameters used for training are the same as those detailed in Annex B and Table 6 of [wav2vec2 paper](https://arxiv.org/pdf/2006.11477.pdf).

## Usage (using the online Inference API)

Just record your voice on the ⚡ Inference API on this webpage, and then click on "Compute", that's all ! 

## Usage (with HuggingSound library)

The model can be used directly using the [HuggingSound](https://github.com/jonatasgrosman/huggingsound) library:

```python
import pandas as pd
from huggingsound import SpeechRecognitionModel

model = SpeechRecognitionModel("Cnam-LMSSC/wav2vec2-italian-phonemizer")
audio_paths = ["./test_rilettura_testo.wav", "./10179_11051_000021.flac"]

# No need for the Audio files to be sampled at 16 kHz here,
# they are automatically resampled by Huggingsound

transcriptions = model.transcribe(audio_paths)

# (Optionnal) Display results in a table :
## transcriptions is list of dicts also containing timestamps and probabilities !

df = pd.DataFrame(transcriptions)
df['Audio file'] = pd.DataFrame(audio_paths)
df.set_index('Audio file', inplace=True)
df[['transcription']]
```

**Output** : 

| **Audio file**                 | **Phonetic transcription (IPA)**                                                                                                                                           |
|:---------------------------|:--------------------------------------------------------------------------------------------------------------------------------------------------------|
| ./test_rilettura_testo.wav | prezɪ lɪ kwatːrotʃɛnto fjorinɪ d̪iː ɔro e reze le debite ɡratsje al pretore sɪ parti e messɔzɪ al merkatantare divɛnne wɔmo sadʒːo e dɪ ɡran manedʒːo    |
| ./10179_11051_000021.flac  | la bʊɔna femina ke ɛra fʊdʒːita il tutːo vedɛva e molto sʊspeza restava e parevale ʊn ora mille annɪ dɪ fʊrarla e dɪ potɛr operare tal effɛtːo          |

## Inference script (if you do not want to use the huggingsound library) : 

```python
import torch
from transformers import AutoModelForCTC, Wav2Vec2Processor
from datasets import load_dataset
import soundfile as sf # Or Librosa if you prefer to ... 

MODEL_ID = "Cnam-LMSSC/wav2vec2-italian-phonemizer"

model = AutoModelForCTC.from_pretrained(MODEL_ID)
processor = Wav2Vec2Processor.from_pretrained(MODEL_ID)

audio = sf.read('example.wav')
# Make sure you have a 16 kHz sampled audio file, or resample it !

inputs = processor(np.array(audio[0]),sampling_rate=16_000., return_tensors="pt")

with torch.no_grad():
  logits = model(**inputs).logits

predicted_ids = torch.argmax(logits,dim = -1)
transcription = processor.batch_decode(predicted_ids)

print("Phonetic transcription : ", transcription)
```

**Output** : 

'ˈsoːno ˈmolto ˈljɛːto di prezenˈtarvi la ˈnɔstra soluˈttsjone per fonemiˈddzaːre fatʃilˈmente ʎi ˈawdjo funˈtsjoːna davˈveːro ˈmolto ˈbɛːne'

## Test Results:

In the table below, we report the Phoneme Error Rate (PER) of the model on Multilingual Librispeech (using the Italian configs for the dataset of course) : 

| Model | Test Set  | PER |
| ------------- | ------------- | ------------- |
| Cnam-LMSSC/wav2vec2-italian-phonemizer | Multilingual Librispeech (Italian) | **4.34%** |


## Citation
If you use this finetuned model for any publication, please use this to cite our work :

```bibtex
@misc {lmssc-wav2vec2-base-phonemizer-italian_2026,
	author       = { Olivier, Malo },
	title        = { wav2vec2-italian-phonemizer (Revision 4d8a3a1) },
	year         = 2026,
	url          = { https://huggingface.co/Cnam-LMSSC/wav2vec2-italian-phonemizer },
	doi          = { 10.57967/hf/7982 },
	publisher    = { Hugging Face }
}
```