EKHOS AI

Captures audio input from microphone or system audio in real-time, processes it through a speech-to-text engine (likely using streaming ASR models), and outputs transcribed text with minimal latency. The architecture appears to use buffered audio chunks fed to an ASR model that maintains state across frames, enabling continuous transcription without waiting for full audio completion.

Accepts pre-recorded audio files (MP3, WAV, M4A, etc.) and video files (MP4, MOV, etc.), extracts audio tracks, and processes them through a speech-to-text model to produce full transcripts. The system likely uses a job queue or async processing pipeline to handle variable file sizes and durations without blocking the UI.

Analyzes generated transcripts using NLP/LLM techniques to identify and suggest corrections for common speech-to-text errors (homophones, context-based word substitutions, punctuation, capitalization). The system likely uses a combination of language models, grammar checkers, and domain-specific correction rules to flag errors and propose fixes without requiring manual review of every word.

Handles diverse audio input formats (MP3, WAV, FLAC, OGG, M4A, etc.) by detecting codec, decoding to a normalized PCM format, and resampling to the target sample rate required by the ASR model. This typically involves FFmpeg or similar codec libraries to abstract format complexity and ensure consistent input to the transcription engine regardless of source format.

Detects speaker changes in audio and labels transcript segments with speaker identities (Speaker 1, Speaker 2, etc.) or names if provided. The system likely uses voice embedding models to cluster similar voices and segment boundaries where speaker changes occur, enabling multi-speaker transcript organization without manual annotation.

Exports finalized transcripts in multiple formats (TXT, PDF, SRT, VTT, DOCX, JSON) with optional metadata (timestamps, speaker labels, confidence scores). The system likely uses templating or format-specific serialization libraries to convert the internal transcript representation into each target format while preserving structure and metadata.

Links transcript text to audio timestamps, enabling users to click on any transcript segment to jump to that point in the audio playback. The system maintains a mapping between text segments and their corresponding audio timestamps, allowing bidirectional navigation (text→audio and audio→text) and precise editing of specific segments without affecting the entire transcript.

Indexes transcript text using full-text search techniques (inverted indexes, tokenization, stemming) to enable fast keyword search across single or multiple transcripts. The system likely builds an in-memory or persistent index of transcript content, allowing sub-second search results even on large transcript collections without scanning every character.