metadata

language:
  - tel
license: apache-2.0
base_model: openai/whisper-large-v3
tags:
  - generated_from_trainer
datasets:
  - jayasuryajsk/google-fleurs-te-romanized
model-index:
  - name: Wishper-Large-V3-spoken_telugu_romanized
    results: []

Wishper Large V3 - Romanized Spoken Telugu

This model is a fine-tuned version of openai/whisper-large-v3 on the Telugu Romanized 1.0 dataset. It achieves the following results on the evaluation set:

eval_loss: 1.5009
eval_wer: 68.1275
eval_runtime: 591.6137
eval_samples_per_second: 0.798
eval_steps_per_second: 0.1
epoch: 8.6207
step: 1000

Model description

The model is trained to transcipt Telugu conversations in Romanized script, that most people uses in day to day life.

Intended uses & limitations

Limitations: Sometimes, it translates the audio to english directly. Working on this to fix it.

Training and evaluation data

Gpt 4 api was used to convert google-fleurs telugu labels to romanized script. I used english tokenizer, since the script is in english alphabet to train the model.

Usage

from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
import torch

device = "cuda" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32

model_id = "jayasuryajsk/whisper-large-v3-Telugu-Romanized"

model = AutoModelForSpeechSeq2Seq.from_pretrained(
    model_id, torch_dtype=torch_dtype
)
model.to(device)

processor = AutoProcessor.from_pretrained(model_id)

pipe = pipeline(
    "automatic-speech-recognition",
    model=model,
    tokenizer=processor.tokenizer,
    feature_extractor=processor.feature_extractor,
    max_new_tokens=128,
    chunk_length_s=30,
    batch_size=16,
    return_timestamps=True,
    torch_dtype=torch_dtype,
    device=device,
)
result = pipe("recording.mp3", generate_kwargs={"language": "english"})
print(result["text"])

Try this on https://colab.research.google.com/drive/1KxWSaxZThv8PE4mDoLfJv0O7L-5hQ1lE?usp=sharing

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 1e-05
train_batch_size: 20
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
training_steps: 2000
mixed_precision_training: Native AMP

Framework versions

Transformers 4.40.1
Pytorch 2.2.0+cu121
Datasets 2.19.1
Tokenizers 0.19.1