Speech Recognition (ASR) Overview & Comparison Matrix

Speech Recognition (ASR, Automatic Speech Recognition) is the foundational cornerstone of DuRT. In DuRT 3.0, speech recognition is strictly partitioned into two dedicated capability panels: Realtime Speech and File Transcription.

Understanding the core distinction and engine capabilities behind each panel ensures optimal transcription accuracy, latency, and privacy for your workflows.

Core Architecture: Streaming vs Non-Streaming Pipelines

DuRT separates processing pipelines based on the nature of the audio input:

mermaid

flowchart TD
    subgraph AudioSources["Input Audio Sources"]
        S1["Live Audio Stream (audioStream)<br>Microphone / ScreenCaptureKit System Audio"]
        S2["Media Files (audioFile / videoFile)<br>Local Files / Online Streaming URLs"]
    end

    subgraph RealtimePanel["【Panel 1】Realtime Speech"]
        R1["🍎 Apple Native Speech (100% Offline / Zero Network / Free)"]
        R2["⚡️ OneASR Gateway (Low-Latency WebSocket Streaming)"]
        R3["🌐 OpenAI Realtime Channel (Cloud Streaming Transcription)"]
    end

    subgraph FilePanel["【Panel 2】File Transcription"]
        F1["⚡️ OneASR Gateway (Long Media Batch / Web URL Ingest / Millisecond Timestamps / Diarization)"]
        F2["🌐 OpenAI ASR (Whisper-1 Timestamps / GPT-4o-Transcribe)"]
        F3["⚠️ Notice: Apple Native Engine DOES NOT support media file transcription"]
    end

    S1 --> RealtimePanel
    S2 --> FilePanel

IMPORTANT

Key Architectural Constraint: Apple Native Speech Does Not Support File Transcription macOS's built-in speech recognition framework is strictly designed for live microphone dictation and real-time audio streams. In DuRT, Apple Native Speech is exclusively mounted under the Realtime Speech panel and cannot transcribe imported local media files. For file transcription, please use OneASR Gateway or OpenAI ASR.

ASR Engines Feature & Capability Comparison Matrix

Engine Service	App Settings Panel	Realtime Streaming (audioStream)	Media Files (audio/videoFile)	Word-Level Timestamps (Millisecond)	Speaker Diarization	Online Video URL Ingest	Run Mode	Pricing & Cost	Best For
🍎 Apple Native Speech	Realtime Speech (Only)	✅ Supported (Internal / Mic)	❌ Not Supported	❌ Sentence level	❌ Not supported	❌ Not supported	100% On-Device Offline	Permanently Free, 0 data usage	Online meetings, live YouTube dual captions, private voice dictation
⚡️ OneASR Gateway	Realtime & File (Dual)	✅ Supported (WebSocket)	✅ Supported (Long media batch)	✅ Supported (Karaoke alignment)	✅ Supported (Speaker 0/1)	✅ Supported (Bilibili/YouTube URLs)	Self-Hosted / LAN GPU	Open Source Free	Team GPU clusters, video creators extracting subtitles from web URLs
🌐 OpenAI ASR	Realtime & File (Dual)	✅ Supported (Live channel)	✅ Supported (Whisper/4o)	✅ Supported (Whisper-1 aligned)	❌ Post-processing needed	❌ Local upload required	Cloud API	Official Token Rates (BYOK)	Academic lectures, legal interviews, multilingual mixed dialogue

Capability Panels Breakdown

1. Realtime Speech

Input Data Contract: audioStream (PCM continuous audio buffer);
Operational Profile: Real-time consumption with sub-500ms latency, powering desktop floating dual subtitles and live mic dictation.
Recommendations:
- Primary: Apple Native Speech (zero network dependency, zero cost, energy-efficient on Apple Silicon);
- Secondary: OneASR Streaming (specialized multilingual models or custom acoustic environments).

2. File Transcription

Input Data Contract: audioFile / videoFile (local file paths or remote direct media URLs);
Operational Profile: VAD-segmented chunking and high-concurrency offline transcription, producing millisecond-precision word timestamps and speaker labels.
Recommendations:
- Primary: OneASR (paste streaming URLs directly without manual downloading; millisecond word timestamps & speaker diarization);
- Highest Accuracy: OpenAI Whisper-1 (exceptional robustness in noisy environments).

Speech Recognition (ASR) Overview & Comparison Matrix ​

Core Architecture: Streaming vs Non-Streaming Pipelines ​

ASR Engines Feature & Capability Comparison Matrix ​

Capability Panels Breakdown ​

1. Realtime Speech ​

2. File Transcription ​