Stefania Budano
Distributed ASR: split model inference for multi-device implementation.
Rel. Daniele Apiletti, Domenico Actis Grosso. Politecnico di Torino, Corso di laurea magistrale in Data Science And Engineering, 2026
|
|
PDF (Tesi_di_laurea)
- Tesi
Accesso limitato a: Solo utenti staff fino al 2 Ottobre 2027 (data di embargo). Licenza: Creative Commons Attribution Non-commercial No Derivatives. Download (4MB) |
Abstract
The field of Automatic Speech Recognition (ASR) in ML has seen great improvement in recent years, with the development of new models that aim to improve the transcription accuracy, often at the cost of being bigger and more computationally expensive. Driven by the usefulness of these models in potentially improving the quality of life for people suffering from low-hearing, this work focuses on optimizing a pre-existing model to allow for inference on a resource-constrained offline device, such as a smart intercom system. For this purpose, after briefly analyzing the state of the art, the OpenAI Whisper tiny model was selected, and, based on the architecture of an intercom system in a multiple device setting, the model execution was split in three parts, simulating the execution on three physical devices, two of which are an active part of the call, and a third one which is borrowed for additional computation capability.
The model execution is as such split into encoder, decoder, and tokenizer, with the former two components also quantized to float16
Relatori
Anno Accademico
Tipo di pubblicazione
Numero di pagine
Corso di laurea
Classe di laurea
Aziende collaboratrici
URI
![]() |
Modifica (riservato agli operatori) |
