README.md

---
language:
- ro
license: apache-2.0
tags:
- whisper-event
datasets:
- mozilla-foundation/common_voice_11_0
- gigant/romanian_speech_synthesis_0_8_1
model-index:
- name: Whisper Medium Romanian
  results:
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: mozilla-foundation/common_voice_11_0 ro
      type: mozilla-foundation/common_voice_11_0
      config: ro
      split: test
      args: ro
    metrics:
    - name: Wer
      type: wer
      value: 4.73
  - task:
      name: Automatic Speech Recognition
      type: automatic-speech-recognition
    dataset:
      name: google/fleurs ro
      type: google/fleurs
      config: ro
      split: test
      args: ro
    metrics:
    - name: Wer
      type: wer
      value: 19.64
metrics:
- wer
---

# Whisper Medium Romanian

This model is a fine-tuned version of [openai/whisper-medium](https://huggingface.co/openai/whisper-medium) on the Common Voice 11.0 dataset, and the Romanian speech synthesis corpus.
It achieves the following results on the evaluation set:
- eval_loss: 0.06453
- eval_wer: 4.717
- epoch: 7.03
- step: 3500

## Model description

The architecture is the same as [openai/whisper-medium](https://huggingface.co/openai/whisper-medium).

## Training and evaluation data

The model was trained on the Common Voice 11.0 dataset (`train+validation+other` splits) and the Romanian speech synthesis corpus, and was tested on the `test` split of the Common Voice 11.0 dataset.

## Usage
Inference with 🤗 Pipeline
```python
import torch
from datasets import load_dataset
from transformers import pipeline

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
# Load pipeline
pipe = pipeline("automatic-speech-recognition", model="gigant/whisper-medium-romanian", device=device)
# NB: set forced_decoder_ids for generation utils
pipe.model.config.forced_decoder_ids = pipe.tokenizer.get_decoder_prompt_ids(language="ro", task="transcribe")

# Load data
ds_mcv_test = load_dataset("mozilla-foundation/common_voice_11_0", "ro", split="test", streaming=True)
test_segment = next(iter(ds_mcv_test))
waveform = test_segment["audio"]

# NB: decoding option
# limit the maximum number of generated tokens to 225
pipe.model.config.max_length = 225 + 1
# sampling
# pipe.model.config.do_sample = True
# beam search
# pipe.model.config.num_beams = 5
# return
# pipe.model.config.return_dict_in_generate = True
# pipe.model.config.output_scores = True
# pipe.model.config.num_return_sequences = 5
# Run
generated_sentences = pipe(waveform)["text"]
```
Inference with 🤗 low-level APIs
```python
import torch
import torchaudio

from datasets import load_dataset
from transformers import AutoProcessor, AutoModelForSpeechSeq2Seq
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# Load model
model = AutoModelForSpeechSeq2Seq.from_pretrained("gigant/whisper-medium-romanian").to(device)
processor = AutoProcessor.from_pretrained("gigant/whisper-medium-romanian", language="romanian", task="transcribe")

# NB: set forced_decoder_ids for generation utils
model.config.forced_decoder_ids = processor.get_decoder_prompt_ids(language="ro", task="transcribe")
# 16_000
model_sample_rate = processor.feature_extractor.sampling_rate

# Load data
ds_mcv_test = load_dataset("mozilla-foundation/common_voice_11_0", "ro", split="test", streaming=True)
test_segment = next(iter(ds_mcv_test))
waveform = torch.from_numpy(test_segment["audio"]["array"])
sample_rate = test_segment["audio"]["sampling_rate"]
# Resample
if sample_rate != model_sample_rate:
    resampler = torchaudio.transforms.Resample(sample_rate, model_sample_rate)
    waveform = resampler(waveform)
# Get feat
inputs = processor(waveform, sampling_rate=model_sample_rate, return_tensors="pt")
input_features = inputs.input_features
input_features = input_features.to(device)

# Generate
generated_ids = model.generate(inputs=input_features, max_new_tokens=225)  # greedy
# generated_ids = model.generate(inputs=input_features, max_new_tokens=225, num_beams=5)  # beam search
# Detokenize
generated_sentences = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]

# Normalise predicted sentences if necessary
```

The code was adapted from [bofenghuang/deprecated-whisper-large-v2-cv11-french-punct-plus](https://huggingface.co/bofenghuang/deprecated-whisper-large-v2-cv11-french-punct-plus).

## Training procedure

### Training hyperparameters

The following hyperparameters were used during training:
- learning_rate: 1e-05
- train_batch_size: 32
- eval_batch_size: 32
- seed: 42
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: linear
- lr_scheduler_warmup_steps: 500
- training_steps: 5000
- mixed_precision_training: Native AMP

### Framework versions

- Transformers 4.26.0.dev0
- Pytorch 1.13.0+cu117
- Datasets 2.7.1.dev0
- Tokenizers 0.13.2