Audio AI Field Guide
A bilingual, mental-model-first reference on audio machine learning, covering datasets, transformer architectures, speech recognition, text-to-speech, and production voice pipelines.
Audio AI Field Guide is a reference guide to modern speech and audio machine learning, organized around explaining the design decisions behind common architectures rather than only describing how a specific model works. It maps each unit to a core mental model, tracing a common five-stage pipeline shared by most audio systems: raw signal, preprocessing, feature extraction, a transformer backbone, and a task head producing a prediction, and draws an explicit parallel to the same pipeline in NLP transformers such as BERT and GPT.
Its units cover working with audio data, the landscape of audio AI tasks, how audio transformers work, automatic speech recognition including alignment methods such as CTC and sequence-to-sequence models, evaluating and fine-tuning ASR systems, text-to-speech datasets and pretrained models, and a walkthrough of fine-tuning SpeechT5. It closes with a systems-integration unit, comparison tables, and a glossary of more than thirty terms.
It is available in both English and a Persian version, and is licensed under the MIT license.