MUMBAI, India, Nov. 2 -- Intellectual Property India has published a patent application (202617092741 A) filed by Google LLC on July 30, 2026, for Using Synthetic Data To Improve Word Error Rate Of Differentially Private Asr Models.
Inventors include Liu, Hongbin; Shejwalkar, Virat Vishnu; Wang, Lun; Thakkar, Om Dipakbhai; Thakurta, Abhradeep Guha; and Narayanan, Arun.
The application for the patent was published on October 30, 2026, under issue no. 44/2026.
Abstract: A method (600) includes pre-training an audio encoder (210) on a public training utterance set and from a corpus of text utterances (501), sampling a predetermined number of most frequent words (501). The method also includes randomly generating a predetermined number of transcripts (520), and for each corresponding transcripts, processing, using a TTS system (550), the corresponding transcript to generate a corresponding synthetic speech utterance (532). The corresponding transcript and the corresponding synthetic speech utterance form a corresponding synthetic training sample (502). During a first fine-tuning stage, the method also includes fine-tuning an ASR model (200) on the synthetic training samples. During a second fine-tuning stage, the method also includes fine-tuning, using a differentially private parameter-efficient-fine-tuning (DP-PEFT) technique, the ASR model on a plurality of private training samples (402), wherein the DP-PEFT technique updates only a subset of newly added or existing parameters of the pre-trained audio encoder.
Disclaimer: Curated by HT Syndication.
