Skip to content
FloWords

Model Specifications

Detailed technical specifications for the transcription models included in FloWords.

Local Models Overview

Section titled “Local Models Overview”

FloWords includes three local transcription engines. All run entirely on your Mac and are multilingual.

ModelEngineDownloadRAMWERLatency
Whisper Turbo (default)OpenAI • Q5_0~547 MB~2 GB~7-8%~200-800 ms
Parakeet V3NVIDIA • INT8~640 MB~2 GB~6.34%~50-200 ms
Apple SpeechNative macOSNoneMinimal~8%~100-500 ms

Model Details

Section titled “Model Details”

Whisper Turbo

Section titled “Whisper Turbo”
Name: Whisper Turbo
Engine: OpenAI Whisper (Large v3 Turbo, Q5_0)
Backend: whisper.cpp (optimized for Apple Silicon)
Download: ~547 MB
Memory Usage: ~2 GB
Accuracy: ~7-8% WER
Latency: ~200-800 ms
Languages: Multilingual

Best for:

  • General daily use
  • Best balance of accuracy and speed
  • Recommended default engine

Parakeet V3

Section titled “Parakeet V3”
Name: Parakeet V3
Engine: NVIDIA Parakeet (via FluidAudio)
Quantization: INT8
Download: ~640 MB
Memory Usage: ~2 GB
Accuracy: ~6.34% WER
Latency: ~50-200 ms
Languages: Multilingual (English + European)

Best for:

  • Fast, low-latency dictation
  • When speed is the priority
  • Good performance with moderate resources

Apple Speech

Section titled “Apple Speech”
Name: Apple Speech
Engine: Native macOS Speech (SFSpeechRecognizer)
Download: None (built into macOS)
Processing: On-device
Accuracy: ~8% WER (lower than Whisper)
Latency: ~100-500 ms
Languages: Multilingual

Best for:

  • Quick drafts
  • No download required
  • Speed and privacy over precision

Language Support

Section titled “Language Support”
  • Whisper Turbo - multilingual, supports 99+ languages with automatic detection
  • Parakeet V3 - English and European languages
  • Apple Speech - multilingual (Arabic, German, English, Spanish, French, Italian, Japanese, Korean, Portuguese, Chinese, and more)

Model Performance by Hardware

Section titled “Model Performance by Hardware”
Section titled “Apple Silicon (Recommended)”
  • Hardware acceleration for all three models
  • ~2-3x faster than Intel
  • Minimal battery impact
  • All three models run great

Intel Macs

Section titled “Intel Macs”
  • Slower than Apple Silicon, higher CPU usage
  • Parakeet V3 or Apple Speech recommended for better speed

Audio Specifications

Section titled “Audio Specifications”

Input Requirements

Section titled “Input Requirements”
SpecificationValue
Sample Rate16000 Hz
Bit Depth16-bit
ChannelsMono
FormatPCM

FloWords automatically converts audio to these specifications.

Supported Input Formats

Section titled “Supported Input Formats”
FormatExtensionNotes
WAV.wavNative support
MP3.mp3Converted to WAV
M4A.m4aConverted to WAV
AAC.aacConverted to WAV
FLAC.flacConverted to WAV
AIFF.aiffConverted to WAV
CAF.cafConverted to WAV
MP4.mp4Audio extracted
MOV.movAudio extracted

Model Selection Guide

Section titled “Model Selection Guide”

By Use Case

Section titled “By Use Case”
Use CaseRecommended Model
Daily useWhisper Turbo
Maximum speedParakeet V3
No download / quick draftsApple Speech
Highest accuracyWhisper Turbo

By Situation

Section titled “By Situation”
SituationRecommended
Clear speechAny
Low latency neededParakeet V3
Offline, no setupApple Speech
Technical termsWhisper Turbo + dictionary

Next Steps

Section titled “Next Steps”