Fine-Tuned SpeechT5 Model for American English TTS
This model is a fine-tuned version of microsoft/speecht5_tts, trained on a dataset derived from audiobooks recorded by Karen Savage, which are in the public domain on LibriVox.org.
Here are some of the sample audio outputs from this model.
Evaluation
There is currently no standardized metric to evaluate TTS quality effectively.
Evaluation was done mainly through listening tests to determine naturalness.
Live Demo
Click here
- Input numbers as words (e.g., write
10as "ten"). - Punctuation matters—so use proper punctuation to get natural speech.
- Apostrophes: For clarity, shortenings like "you're" must be written fully as "you are".
Usage
import torch
import torchaudio
from transformers import SpeechT5Processor, SpeechT5ForTextToSpeech
from transformers.models.speecht5 import SpeechT5HifiGan
# Load processor, model, and vocoder
processor = SpeechT5Processor.from_pretrained("nambn0321/T5_US_accent_2")
model = SpeechT5ForTextToSpeech.from_pretrained("nambn0321/T5_US_accent_2", use_safetensors=True, trust_remote_code=True)
vocoder = SpeechT5HifiGan.from_pretrained("microsoft/speecht5_hifigan")
# Move to device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = model.to(device)
vocoder = vocoder.to(device)
# For the sake of simplicity, use this predefined speaker embedding
speaker_embedding = torch.tensor([[-7.8568e-02, -4.2079e-03, 1.1993e-02, 1.2876e-02, 3.8205e-03,
-1.9735e-03, -6.8052e-02, -6.2425e-02, 4.2591e-02, 2.0495e-02,
-6.5605e-02, -7.4267e-02, 4.7141e-02, 3.1141e-02, 3.3795e-02,
6.8717e-02, 1.5437e-02, 2.9659e-02, 9.6837e-03, 1.6690e-02,
4.1287e-02, 1.0799e-02, -1.4346e-02, -3.6507e-02, -6.9912e-02,
-1.1495e-02, -5.9190e-02, 5.0997e-03, 3.5220e-02, 2.7239e-02,
-3.0035e-03, 4.0179e-02, 2.7811e-02, -3.7754e-02, 4.2270e-02,
-7.6790e-02, 3.3923e-02, 5.8342e-02, -6.8696e-02, -6.8298e-02,
-1.5029e-03, -5.7018e-02, -4.0267e-03, 5.2543e-02, 1.2046e-02,
-1.1127e-01, -1.9529e-02, 1.1586e-02, -7.0273e-02, 5.7403e-02,
1.9700e-02, 3.5813e-02, 3.8164e-02, 4.1581e-02, -7.9466e-02,
-4.0844e-03, 4.3121e-02, 2.5432e-02, 1.6693e-02, 1.4494e-02,
3.2961e-02, -1.0050e-02, -1.6570e-02, 2.1572e-02, 2.3886e-02,
3.7505e-02, 2.3737e-03, -3.5667e-02, -6.9384e-02, -6.1990e-02,
2.1427e-02, 1.0910e-02, -4.4866e-03, 1.9126e-02, 3.5026e-02,
2.6617e-02, 1.0270e-02, 1.7574e-02, -5.0846e-02, -7.9475e-02,
-5.9455e-02, -5.5634e-02, -5.4523e-02, -6.2594e-02, -3.4710e-02,
-4.8424e-02, -6.5559e-02, 4.3848e-02, -8.9867e-06, 5.7124e-02,
2.9633e-02, -8.8773e-02, 8.2799e-03, -6.3414e-02, 2.7484e-02,
6.6257e-03, 3.2360e-02, 3.4513e-02, -2.0671e-02, -8.1817e-02,
4.1832e-02, -6.9010e-02, -5.7109e-02, 5.1551e-02, 3.6937e-02,
-5.9055e-02, 2.5737e-02, 4.8279e-02, 4.0342e-02, 2.0409e-02,
-7.8760e-02, 4.8960e-02, 6.1605e-02, 1.5055e-03, 4.4753e-02,
5.1425e-02, -6.9668e-02, -3.3952e-02, -5.3081e-02, -3.3253e-02,
2.1449e-02, -7.3866e-02, 1.5239e-02, 3.7210e-02, -7.0857e-02,
4.2094e-02, -7.8425e-02, 2.2612e-02, 4.6070e-02, 3.1248e-02,
2.1681e-02, 9.0710e-03, 2.6234e-02, 3.9768e-02, 2.6416e-02,
-5.9739e-02, -5.3194e-02, 1.1592e-02, -7.3099e-02, -4.0911e-02,
2.9276e-02, 4.0793e-03, -2.7053e-02, 4.3887e-02, -7.4993e-02,
2.8244e-02, 1.4546e-02, -5.5933e-02, 5.4590e-02, -9.8596e-02,
2.3044e-02, -4.3384e-02, -6.2760e-02, 4.9645e-02, 1.9709e-02,
2.2457e-02, 1.0992e-02, -9.1083e-02, -7.2880e-02, 5.3015e-02,
1.4966e-02, 7.6749e-03, 1.2842e-02, -6.0044e-02, 1.4364e-03,
1.2117e-02, 3.7999e-02, 4.1830e-02, 1.7146e-02, 4.1624e-02,
1.9113e-02, -8.6394e-02, 3.9947e-02, -4.5318e-02, -1.5646e-02,
1.7320e-02, -5.8261e-02, 1.3057e-02, 1.7871e-02, -7.2801e-02,
2.7487e-02, -5.1378e-02, 1.0601e-02, 3.2772e-02, -3.3645e-02,
-9.6321e-03, 5.7508e-02, 3.8802e-02, -5.4275e-02, -6.4749e-02,
-2.3990e-02, 4.4422e-02, -5.5291e-02, 2.1329e-02, 3.5870e-02,
1.5788e-02, 1.9083e-02, -2.5848e-03, 3.0792e-02, -2.4433e-02,
4.0921e-02, 2.2340e-02, -4.7077e-02, 5.6612e-03, 2.4069e-02,
1.7687e-02, 5.2614e-02, -1.4121e-02, 4.4471e-02, -4.5358e-02,
3.0660e-03, -8.4165e-02, -4.3935e-02, 5.7635e-02, -4.6062e-02,
2.8475e-02, 2.7438e-02, -7.8207e-02, 3.6834e-02, 3.5305e-02,
-7.9270e-02, 1.5048e-02, -7.7217e-02, -3.3846e-02, 4.0682e-02,
4.5813e-02, 6.3953e-02, 8.8146e-02, 3.9316e-02, 3.6404e-02,
-3.6674e-02, 3.9037e-02, 3.2509e-02, -3.3039e-02, 9.0764e-03,
-1.9967e-02, 3.4478e-02, 2.2831e-02, -6.8772e-04, 5.4448e-02,
-6.7131e-02, 2.6475e-02, -9.6572e-02, 2.7054e-02, -6.1189e-02,
4.2293e-02, 5.5649e-02, 2.4348e-02, 6.6935e-03, 4.2651e-02,
3.7361e-02, 3.3392e-02, 9.3010e-03, -5.7520e-02, 5.3737e-03,
4.5707e-02, 2.8316e-02, -1.5346e-03, -6.4626e-02, 5.0692e-02,
1.4295e-02, -5.4578e-02, 3.8668e-02, 2.1647e-02, 1.4004e-03,
2.3282e-02, 3.1919e-02, 1.2071e-02, 1.3926e-02, -4.4616e-02,
4.2064e-02, -1.8788e-02, 1.6830e-02, -1.6330e-02, -6.7638e-02,
4.5764e-02, 1.6224e-02, 1.3495e-02, -7.7807e-02, -4.8269e-02,
-2.7209e-02, 5.7491e-02, 3.6628e-02, -8.6239e-02, -5.5271e-02,
3.9839e-02, 1.0211e-03, 5.5201e-02, -9.7384e-02, 3.8847e-03,
1.0693e-02, 7.5698e-03, -5.3666e-02, 4.1555e-02, -3.2620e-02,
3.2532e-02, 7.4491e-03, 3.6136e-02, 1.7120e-02, 2.5016e-02,
6.8792e-02, 2.9997e-02, 2.1673e-02, -7.8844e-02, 1.1353e-02,
3.5831e-02, 3.0084e-02, 3.0417e-02, 2.9927e-02, 2.1848e-02,
4.9556e-02, 2.2132e-02, -2.8324e-02, 4.4158e-02, -8.2102e-02,
-6.4570e-02, -2.4734e-02, 3.2701e-02, -7.0163e-02, 5.4873e-02,
-4.7028e-02, 4.4843e-02, -4.5314e-02, 1.0327e-02, 2.8297e-02,
-5.7504e-02, 4.7179e-02, 7.4731e-02, -6.5681e-02, -8.6343e-02,
-6.4412e-02, 3.1260e-02, 1.6076e-02, 4.7171e-02, -7.1781e-02,
4.2377e-02, 3.9755e-02, -3.6226e-02, -7.4231e-03, -6.4577e-02,
3.0569e-02, -5.3078e-02, 2.7852e-02, -7.6148e-03, -7.3751e-02,
2.0000e-02, 2.1321e-02, 1.5519e-02, -3.6516e-02, -5.5269e-02,
-4.3193e-02, -1.7178e-02, -5.1271e-02, 1.0353e-01, 4.1393e-02,
-4.7789e-02, -8.0428e-03, 2.9483e-02, -5.4314e-02, 1.0356e-02,
-1.0647e-01, 2.6810e-02, -1.3466e-02, -9.5602e-04, 5.6365e-02,
-3.4805e-02, -4.8433e-02, 5.5901e-03, 1.0095e-02, 4.4062e-02,
1.3886e-02, 2.7514e-02, -9.5484e-02, 1.4190e-02, -1.3233e-02,
-2.4893e-03, 2.6416e-02, 6.7407e-03, 6.1025e-02, 3.8437e-02,
-7.4136e-02, -1.1276e-01, 1.3998e-02, 4.5844e-02, 1.8342e-02,
-6.7303e-02, 2.9729e-02, -6.0356e-02, 3.4768e-02, 2.6196e-02,
5.8514e-03, 7.3593e-03, -4.2139e-02, 3.0210e-02, 1.5900e-02,
7.0803e-03, 3.3725e-02, -8.8192e-02, 1.3683e-03, 1.4380e-02,
-1.8023e-02, -6.0320e-02, 1.4030e-02, -4.0541e-02, 4.6965e-03,
7.1572e-03, 1.0316e-02, -7.6909e-02, -5.5507e-02, -6.4332e-02,
-6.2764e-02, 2.3172e-02, 1.5215e-02, -1.5576e-02, 2.3396e-02,
-5.4251e-02, 1.7465e-02, -9.1552e-02, -1.4350e-01, -1.5228e-02,
-5.0016e-02, 1.5546e-02, 1.9486e-02, -2.2702e-02, -6.0833e-02,
1.8424e-02, 4.1719e-02, 3.1578e-02, 2.6568e-02, -4.9155e-02,
-5.2004e-02, -1.8590e-02, -2.7371e-02, 3.8227e-02, 3.2638e-02,
7.9873e-03, 4.5671e-02, 2.4781e-02, -6.7724e-02, -7.6685e-02,
1.3213e-02, 1.9150e-02, 2.0911e-02, 4.8548e-03, 5.5948e-02,
2.9883e-02, 2.2585e-02, 1.0647e-02, 9.4530e-03, -1.6939e-02,
4.8591e-02, 2.6256e-02, 4.8367e-02, 5.7640e-02, 1.4820e-02,
1.0206e-02, 2.1576e-02, -6.3301e-02, -6.1438e-02, 4.9681e-02,
-1.4290e-02, 9.2644e-03, 4.7036e-02, 2.7807e-02, -4.7537e-02,
2.8718e-02, 3.9035e-02, -6.9315e-02, 2.0267e-02, 9.3887e-03,
-2.3518e-03, 3.0030e-02, 2.0438e-02, 4.7360e-03, -1.5699e-02,
-7.5235e-02, 1.8405e-02, -5.7478e-03, 2.8843e-02, 4.1911e-02,
-6.1657e-02, -5.3779e-02, 1.2746e-02, 2.4689e-02, 2.3149e-02,
3.2983e-02, -5.4079e-02, 2.3033e-02, -1.2222e-02, -1.3194e-02,
-4.7920e-02, 3.9478e-02, -5.1594e-02, 1.0203e-02, 8.6237e-04,
-1.2024e-03, -5.9529e-02, 1.3870e-02, -6.7391e-02, -7.4410e-02,
9.1564e-03, 2.5374e-02, -8.6928e-02, 3.2397e-02, -4.7997e-02,
-1.4516e-02, -6.2727e-02, 4.8488e-02, 6.5368e-02, -2.2742e-02,
3.6199e-02, -7.3590e-02]]).to(device)
# Input text
text = "Hello, how are you doing?"
# Process text
inputs = processor(text=text, return_tensors="pt").to(device)
# Generate speech
with torch.no_grad():
speech = model.generate_speech(inputs["input_ids"], speaker_embedding, vocoder=vocoder)
# Save to file
if speech.dim() == 1:
speech = speech.unsqueeze(0)
torchaudio.save("output.wav", speech.cpu(), sample_rate=16000)
Nom