How they use Nebius AI Cloud

IIIT-Delhi researchers use Nebius compute through its research credits programme for speech AI. Nebius describes work on voice verification, deepfake detection, speech recognition and translation, with a focus on Indic languages.

At the institute’s Infosys Centre for AI, Nebius supported ArticulateX, a speech-to-speech translation model that represents how people physically produce sounds. Its pipeline combines a Conformer-based speech-to-articulator encoder, a spectrogram generator and a HiFi-GAN vocoder, with optional diffusion post-processing. LoRA compresses the encoder to 3.1 million parameters. The 70-million-parameter model achieved BLEU scores of 37.24 for French-to-English and 20.07 for German-to-English translation. Nebius reports a naturalness score of 4.84 out of 5 and a real-time factor of 0.97. Its research feature describes an experimental prototype covering 11 Indian languages, while noting that longer conversational speech and broader real-world conditions still require evaluation.

A separate IIIT-Delhi paper acknowledges Nebius Research Grant support for speech emotion recognition research. That work benchmarks speech models and investigates selective fine-tuning, feature pooling and loss functions. Its best configuration combined 15 trainable WavLM layers, multi-layer attentive statistics pooling and focal loss, reaching 66.25% development accuracy and an F1 score of 0.6524.