facebook

/

mms-tts-mlg

like 3

Hello,

Amazing! Any information on the training data?
This is the first time I have seen a TTS implemented for Malagasy, and honestly for a first, it's a good one! Intonation is a bit odd, but pronunciation for native words is mostly on point.

Can we have a clear list of requirements in the model documentation?
I managed to run and generate a wav file using these steps:

Script:

import numpy as np
import torch
import wave

from transformers import VitsModel, AutoTokenizer

model = VitsModel.from_pretrained("facebook/mms-tts-mlg")
tokenizer = AutoTokenizer.from_pretrained("facebook/mms-tts-mlg")

with open('text.txt', 'r') as f:
    text = f.read()

inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    output = model(**inputs).waveform


# Assuming `output` is your Torch tensor containing audio data

# Convert Torch tensor to NumPy array
output_np = output.detach().cpu().numpy()

# Ensure data is in the correct shape
if len(output_np.shape) == 1:  # If it's a 1D array, reshape it to have two dimensions
    output_np = output_np.reshape(-1, 1)
elif len(output_np.shape) > 2:  # If it has more than 2 dimensions, raise an error
    raise ValueError("Data has more than 2 dimensions")

# Normalize audio data to the range [-1, 1]
max_val = np.max(np.abs(output_np))
if max_val > 1:
    output_np = output_np / max_val

# Scale the audio data to fit within the range of signed 16-bit integers
scaled_output = np.int16(output_np * 32767)

# Specify the output file path
output_path = "techno.wav"

# Ensure rate is a positive integer
rate = int(model.config.sampling_rate)
if rate <= 0:
    raise ValueError("Sampling rate must be a positive integer")

# Open the WAV file for writing
with wave.open(output_path, 'w') as wf:
    # Set parameters
    wf.setnchannels(1)  # Mono audio
    wf.setsampwidth(2)  # 16-bit encoding
    wf.setframerate(rate)  # Sampling rate

    # Write audio data
    wf.writeframes(scaled_output.tobytes())

Direct Requirements:

torch==2.2.0
Wave==0.0.2
numpy==1.24.4
transformers==4.38.2