Create a benchmark audio dataset for evaluating deepfake detectors

[dennyabrain]

[plon-Susk7]

AI4Bharat Indic-TTS

We'll be using AI4Bharat Indic-TTS as our first model to generate audio.
AI4Bharat Indic-TTS is open-source SOTA Text-To-Speech model for 13 Indian languages: Assamese, Bengali, Bodo, Gujarati, Hindi, Kannada, Malayalam, Manipuri, Marathi, Odia, Rajasthani, Tamil and Telugu.

Github link : https://github.com/AI4Bharat/Indic-TTS

[plon-Susk7]

Massively Multilingual Speech (MMS) [Facebook]

This model is part of Facebook's Massively Multilingual Speech project, aiming to provide speech technology across a diverse range of languages.
This model is fairly new, published in 2023 and supports wide range of languages. The model architecture is based of VITs. More about the model here https://huggingface.co/facebook/mms-tts-hin.

[aatmanvaidya]

@plon-Susk7 just a random question, to generate audio from the Indic TTS model, you are randomly generating text and feeding it to it right? or picking up the text from somewhere? This thing won't matter much, the goal is to generate audio, but I am just asking casually for curiosity.

[plon-Susk7]

@plon-Susk7 just a random question, to generate audio from the Indic TTS model, you are randomly generating text and feeding it to it right? or picking up the text from somewhere? This thing won't matter much, the goal is to generate audio, but I am just asking casually for curiosity.

Generated random monologue using ChatGPT, I'll also include txt file with all prompts in the dataset.

[plon-Susk7]

Coqui-TTS

This model is text to speech capable of cloning voices. https://github.com/coqui-ai/TTS

Architecture:

• Text2Spec models (Tacotron, Tacotron2, Glow-TTS, SpeedySpeech).
• Speaker Encoder to compute speaker embeddings efficiently.
• Vocoder models (MelGAN, Multiband-MelGAN, GAN-TTS, ParallelWaveGAN, WaveGrad, WaveRNN)