We tested Speakeasy’s Core ML whisper engine against cloud APIs across M1, M2, M3, and M4 Macs. Measure exact latency, Word Error Rate (WER), RAM footprint, and battery efficiency.
From key-release to text injection
On Python/TypeScript prompts
100% on-device Neural Engine
Deep sleep when not dictating
| Processor / System | Compute Unit | First-Token Latency | Real-Time Factor (RTF) | Peak Unified RAM |
|---|---|---|---|---|
| Apple M4 Pro / Max | 38-Core Neural Engine | 380 ms | 0.08x (12x faster than realtime) | 920 MB |
| Apple M3 / M3 Pro | 16-Core Neural Engine | 410 ms | 0.11x (9x faster than realtime) | 940 MB |
| Apple M2 / M2 Max | 16-Core Neural Engine | 445 ms | 0.13x (7.5x faster than realtime) | 950 MB |
| Apple M1 (Base 2020) | 16-Core Neural Engine | 490 ms | 0.16x (6x faster than realtime) | 960 MB |
| Intel Core i9 (2019 Mac) | CPU AVX2 Fallback | 1,240 ms | 0.48x (2x faster than realtime) | 1,420 MB |
| Cloud SaaS (Wispr / Aqua) | Remote GPU Cluster | 750 ms – 1,800 ms (RTT-dependent) | Variable (Network latency bound) | 650 MB (Electron app) |
Lower WER indicates higher accuracy. Tested on 500 standardized audio prompts across conversational prose, technical programming syntax (TypeScript/Rust), and noisy café audio.
| Engine / Solution | Conversational Speech WER | Technical Code & CLI WER | Background Noise Resilience |
|---|---|---|---|
| Speakeasy (Core ML Turbo) | 2.1% | 2.4% (With Dev Syntax) | High (Hardware VAD) |
| Wispr Flow (Cloud) | 2.3% | 4.8% (Struggles with flags) | Moderate |
| Apple Built-in Dictation | 5.8% | 16.4% (No code context) | Low |
| Superwhisper (Whisper Base) | 3.4% | 6.1% | Moderate |
Because Speakeasy executes on the dedicated low-power Apple Neural Engine instead of spinning up high-wattage GPU cores, battery drain is virtually undetectable during normal workday usage.
Harness the full power of your Apple Silicon hardware with zero monthly fees.
Download for macOS ($69)