Sindhi-TTS

This model is a fine-tuned version of fahadqazi/Sindhi-TTS on the None dataset. It achieves the following results on the evaluation set:

• eval_loss: 0.4602
• eval_runtime: 47.8291
• eval_samples_per_second: 36.421
• eval_steps_per_second: 18.211
• epoch: 13.2653
• step: 6500

How to use

  from transformers import SpeechT5ForTextToSpeech, SpeechT5ForSpeechToText
  from transformers import SpeechT5Processor
  from transformers import AutoTokenizer
  from transformers import SpeechT5HifiGan
  import torch
  from IPython.display import Audio as IPythonAudio

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

  # imporing speech processor from another repo
  processor = SpeechT5Processor.from_pretrained("Sana1207/Hindi_SpeechT5_finetuned")

  # importing tokenizer and assigning it to the speech processor
  tokenizer = AutoTokenizer.from_pretrained("fahadqazi/Sindhi-TTS")
  processor.tokenizer = tokenizer

  # importing the model
  model = SpeechT5ForTextToSpeech.from_pretrained("fahadqazi/Sindhi-TTS")

  # importing the vocoder from microsoft's repository
  vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan").to(device)

  # loading random vocodings (the voice)
  embeddings_dataset = load_dataset("Matthijs/cmu-arctic-xvectors", split="validation")
  speaker_embeddings = embeddings_dataset[7306]["xvector"]
  speaker_embeddings = torch.tensor(speaker_embeddings).to(device).unsqueeze(0)


  # Generating Speech
  text = "ڪهڙا حال آهن"
  inputs = processor(text=text, return_tensors="pt").to(device)


  speech = model.generate_speech(inputs["input_ids"], speaker_embeddings, vocoder=vocoder)

  IPythonAudio(speech.cpu().numpy(), rate=16000, autoplay=True)

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: 0.0001
• train_batch_size: 16
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 32
• optimizer: Use adamw_torch with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 200
• training_steps: 10000
• mixed_precision_training: Native AMP

Framework versions

• Transformers 4.46.2
• Pytorch 2.5.1+cu121
• Datasets 3.1.0
• Tokenizers 0.20.3