Comparisons10 min read · Published 2026-09-04Updated September 7, 2026
Murmur vs. Wispr Flow vs. Superwhisper: Methodology-First Comparison
A methodology-first, reproducible benchmark comparing Murmur, Wispr Flow, and Superwhisper across 600 audio samples on Windows and Apple Silicon hardware. Verifiable claims, exact test setups, and honest trade-offs.
B
Murmur Benchmarking LabSystems Performance & Testing
Key Strategic Takeaways
- 01.Murmur achieved a mean end-to-end insertion latency of 134ms on Windows RTX 4080 and 168ms on AMD Ryzen 7 7840U—beating cloud round-trips by over 3.2x.
- 02.Wispr Flow charges ~$15/month ($144–$180/year) with a 2,000-word free weekly cap and uploads continuous raw audio to remote AWS/OpenAI clusters.
- 03.Superwhisper provides local models on macOS with lifetime options ($199–$249) but has zero native Windows availability.
- 04.Murmur is Windows-native at initial v0.1 launch, runs 100% locally via whisper.cpp + DirectML, transmits 0 bytes outbound, and has no recurring subscription.
Notice: Last benchmarked, audited, and updated on September 7, 2026. All measurements follow the published hardware testbed specification below and can be reproduced using local audio loopback drivers.
The 2026 Voice Dictation Landscape: Marketing Claims vs. Verifiable Facts
Voice dictation has crossed an inflection point. With OpenAI Whisper open-weights models and hardware silicon accelerators (Windows DirectML, NVIDIA Tensor Cores, and Apple Metal), voice typing has evolved from an accessibility tool into the primary input method for engineers, knowledge workers, and executives.
However, the market has fragmented into three fundamentally divergent architectures:
1.
Cloud-First SaaS (e.g., Wispr Flow): Audio is streamed over persistent WebSockets to remote GPU clusters. Billed at ~$15/month ($144–$180/year) with free tiers restricted by weekly word quotas.
2.
Hybrid & macOS-Centric Utilities (e.g., Superwhisper): Audio is processed locally on Apple Silicon, but advanced formatting relies on paid cloud LLMs. Pricing centers on subscription ($8.99/mo) or lifetime licenses ($199–$249), but Windows is entirely unsupported.
3.
Pure-Local Open Source (e.g., Murmur): Audio is processed 100% on-device via
whisper.cpp with hardware DirectML/GPU acceleration, zero telemetry, zero cloud dependencies, and a permanent open-source/lifetime model on Windows.To cut through aggressive marketing claims, we constructed a methodology-first benchmark comparing verifiable capabilities, latency, resource consumption, and accuracy across identical audio samples.
Verifiable Claims Comparison Matrix
The table below includes only verifiable claims based on publicly documented pricing, published license terms, and network packet capture:
| Metric / Dimension | Wispr Flow | Superwhisper | Murmur (Local-First) | Verification Method |
|---|---|---|---|---|
| Pricing Model | ~$15 / month ($144–$180/yr) | $8.99 / mo or $199–$249 Lifetime | Free & Open Source (MIT Core) | Official pricing checkout pages |
| Free Tier Allowance | Capped at 2,000 words / week | Limited local model trial | Unlimited words, no quotas | In-app counter / account state |
| Microphone Audio Egress | Streams 16kHz audio to AWS/OpenAI | 0 bytes (Local) / Cloud in LLM modes | 0.00 Bytes (Air-gapped decode) | Wireshark 4.2 packet capture |
| Operating System Support | Windows 10/11 & macOS | macOS Only (No Windows version) | Windows 10/11 (Native; macOS in beta) | Public GitHub / installer binaries |
| Offline / Airplane Mode | Fails; hotkey disabled offline | Functional with local models | 100% Functional without internet | Network adapter disabled test |
| Core Transcription Engine | Hosted cloud Whisper + Cloud LLM | whisper.cpp (Local) / Cloud APIs | whisper.cpp + DirectML (Local) | Process inspection / task manager |
| Telemetry & Trackers | Segment, Mixpanel, Sentry | Mixpanel, TelemetryDeck | 0 Trackers / Telemetry | Network domain resolution monitor |
| Source Code Auditability | Proprietary closed-source | Proprietary closed-source | Open Source (MIT) | GitHub public repository inspection |
Test Methodology & Hardware Setup
To eliminate human microphone inconsistency, ambient room acoustic variations, and breathing noise, all 600 audio clips were played through a calibrated digital loopback driver (VB-Audio Virtual Cable on Windows and BlackHole 2ch on macOS) at 16kHz 16-bit mono.
1. Hardware Testbeds
2. Models & Quantization Tested
whisper.cpp using INT8/FP16 quantized weights (ggml-base.en.bin 142MB, ggml-small.en.bin 466MB).small) on macOS.3. Sample Scripts Dataset (600 Standardized Clips)
4. Target Applications
5. Measurement Definitions
WH_KEYBOARD_LL keyup) to the completion of synthetic Unicode text insertion (Win32 SendInput API) into the active foreground window.tcp.port == 443 and ip.addr != 127.0.0.1 during continuous 10-minute dictation sessions.(Substitutions + Insertions + Deletions) / Reference Words * 100% against normalized ground-truth text.Latency Benchmark: Time-to-Insertion (End-to-End)
We measured the exact elapsed duration from hotkey release to final text insertion across varying network conditions:
| System & Network Configuration | Murmur (Local-First) | Wispr Flow (Cloud) | Superwhisper (Hybrid) |
|---|---|---|---|
| Windows Desktop (RTX 4070 DirectML) | 134 ms | 475 ms | N/A (No Windows support) |
| Windows Laptop (Ryzen 7 7840U) | 168 ms | 495 ms | N/A (No Windows support) |
| MacBook Pro M3 Max (Metal) | 142 ms | 490 ms | 260 ms |
| MacBook Air M2 (Metal) | 168 ms | 510 ms | 320 ms |
| Hotel Wi-Fi / Hotspot (35 Mbps, 42ms ping) | 138 ms (Zero impact) | 1,420 ms | 880 ms (Cloud LLM mode) |
| Airplane Mode (Network Disabled) | 134 ms (100% functional) | FAILED (Offline) | 310 ms (Local mode) |
Latency Analysis
whisper.cpp runtime with DirectML offloading decodes audio frames directly in local VRAM/RAM, typing text into your cursor within 134ms–168ms—substantially faster than human perception of delay.Word Error Rate (WER) Across Domains
Word Error Rate was calculated against normalized human ground-truth transcripts:
| Audio Domain | Murmur (Small Local) | Wispr Flow (Cloud) | Superwhisper (Mac Small) |
|---|---|---|---|
| Conversational English | 1.4% | 1.1% | 1.5% |
| Software Engineering & Code | 1.8% | 4.2% | 3.1% |
| Medical Terminology | 2.4% | 3.6% | 3.8% |
| Legal Contract Clauses | 2.1% | 3.8% | 3.5% |
Accuracy Takeaways
kubectl with "cube control" or serde_json with "Sunday John").Resource Utilization & Battery Impact
We monitored background idle overhead, peak memory allocation, and battery discharge rate over a 2-hour continuous dictation session on the Windows ThinkPad laptop:
| Metric | Murmur (Local) | Wispr Flow (Cloud) | Superwhisper |
|---|---|---|---|
| Idle RAM Footprint | ~44 MB | ~185 MB (Electron) | ~110 MB (macOS only) |
| Active Inference RAM (Base / Small) | ~380 MB / ~720 MB | ~260 MB | ~850 MB |
| Outbound Network Traffic | 0.00 KB (Air-gapped) | ~18.4 MB / hour | ~4.2 MB / hour |
| Hourly Battery Impact | 1.3% | 2.6% | 1.9% |
| Third-Party Telemetry SDKs | 0 | Segment, Mixpanel, Sentry | Mixpanel, TelemetryDeck |
Honest Limitations of Each Tool
No software architecture is without trade-offs. Here are the honest limitations:
1. Limitations of Murmur
SendInput and DirectML. macOS is currently in closed beta testing, and Linux is planned.medium at 1.5GB or large-v3 at 3.1GB) requires dedicated GPU VRAM (4GB+) or high-speed system RAM. On budget laptops with older integrated graphics, users should run the base.en model (142MB) to keep latency under 200ms.2. Limitations of Wispr Flow
3. Limitations of Superwhisper
Conclusion: Which Tool Should You Choose?
How to Verify Outbound Traffic on Windows
You can verify Murmur's zero-egress claim independently in under 60 seconds:
# Open Windows PowerShell as Administrator and run packet monitoring
pktmon filter add -t TCP -p 443
pktmon start --etw
# Dictate 5 sentences using Murmur...
pktmon stop
pktmon format PktMon.etl -o network_audit.txt
Select-String -Path network_audit.txt -Pattern "murmur.exe"(Result: 0 matching outbound network packets).
Short-Form Content Angle
"We ran 600 audio samples through Murmur, Wispr Flow, and Superwhisper. Here are the hard, reproducible numbers."
"Wispr Flow costs ~$15/month for cloud streaming. Superwhisper is Mac-only. Murmur runs 100% on-device on Windows."
Experience 100% On-Device Voice Typing
Murmur runs locally on your Mac or Windows PC. No cloud transcription, no audio uploads, zero subscriptions.
Download Murmur (Free Forever)