What is Automatic Speech Recognition (ASR)?

Automatic Speech Recognition (ASR)—often referred to as speech-to-text (STT)—is the algorithmic process of converting continuous human vocal sound waves into written linguistic text.

From its origins at Bell Labs in the 1950s (with the Audrey digit recognizer) to modern transformer-based neural architectures, ASR has evolved from fragile laboratory experiments into an indispensable tool for business transcription, accessibility, and human-computer interfaces.

The mechanical pipeline of modern ASR

Modern neural ASR systems process spoken language through four sequential stages:

1. Acoustic preprocessing and feature extraction

Continuous analog audio captured from a microphone or system loopback is sampled at a standard frequency (typically 16 kHz, 16-bit PCM). The raw waveform is converted into a log-Mel spectrogram using a Short-Time Fourier Transform (STFT). This 2D visual representation maps frequency distributions over time, matching the human cochlea’s perceptual frequency response.

2. Acoustic encoder representation

An encoder neural network (such as a Conformer or Transformer) processes the log-Mel spectrogram patches. It extracts rich acoustic embeddings that capture phonemic boundaries, tone inflections, and speech cadences while filtering out background ambient noise.

3. Autoregressive linguistic decoding

A decoder network translates the acoustic embeddings into subword text tokens (such as Byte-Pair Encoding or SentencePiece tokens). Modern models condition token generation on both preceding acoustic context and language probability distributions.

4. Text post-processing and punctuation

The raw token stream is detokenized into readable prose, applying capitalized letters, periods, commas, and numeric formatting.

Cloud ASR vs On-Device ASR

Performance Metric	Cloud-Hosted ASR (AWS, Google, Deepgram)	On-Device ASR (Oyma, Whisper)
Network Dependency	Mandatory high-bandwidth internet	Zero; 100% offline functionality
Privacy Risk	Audio streams sent to multi-tenant servers	Audio never leaves client device memory
Latency Variance	Subject to network jitter and queues	Deterministic sub-second execution
Operating Cost	Recurring per-minute API fees	One-time local hardware compute

ASR implementation in Oyma on macOS

In Oyma, ASR is engineered exclusively for local execution on Apple Silicon hardware:

• Incoming call audio is streamed directly from macOS ScreenCaptureKit into memory buffers.
• The audio passes into an embedded Whisper model compiled with Metal and Core ML optimizations.
• Text tokens render in real time alongside your meeting notes.
• Audio and transcripts remain completely private on your local Mac disk, satisfying strict organizational governance guidelines.