Model Specifications

Detailed technical specifications for all transcription models supported by FloWords.

---

Whisper Models Overview

Whisper is OpenAI’s automatic speech recognition (ASR) system. FloWords uses whisper.cpp, an optimized C++ implementation for Apple Silicon.

Model Comparison

Model	Parameters	Size	VRAM	English WER	Speed Factor
Tiny	39M	75 MB	~1 GB	8.4%	~32x
Tiny.en	39M	75 MB	~1 GB	7.5%	~32x
Base	74M	142 MB	~1.5 GB	5.0%	~16x
Base.en	74M	142 MB	~1.5 GB	4.3%	~16x
Small	244M	466 MB	~2 GB	3.4%	~6x
Small.en	244M	466 MB	~2 GB	3.0%	~6x
Medium	769M	1.5 GB	~5 GB	2.5%	~2x
Medium.en	769M	1.5 GB	~5 GB	2.1%	~2x
Large-v3	1550M	3 GB	~10 GB	2.0%	~1x

Note

WER = Word Error Rate (lower is better). Speed Factor is relative to Large model (higher is faster).

---

Whisper Model Details

Tiny

Name: Whisper Tiny
Parameters: 39 million
File Size: 75 MB
Memory Usage: ~1 GB
Layers: 4 encoder, 4 decoder
Dimension: 384
Heads: 6

Best for:

• Quick notes
• Low-memory systems
• Fastest transcription

Trade-offs:

• Lower accuracy
• May struggle with accents
• Limited noise handling

---

Base

Name: Whisper Base
Parameters: 74 million
File Size: 142 MB
Memory Usage: ~1.5 GB
Layers: 6 encoder, 6 decoder
Dimension: 512
Heads: 8

Best for:

• General daily use
• Good balance of speed and accuracy
• Most Mac configurations

Trade-offs:

• Moderate accuracy
• Some errors with technical terms

---

Small

Name: Whisper Small
Parameters: 244 million
File Size: 466 MB
Memory Usage: ~2 GB
Layers: 12 encoder, 12 decoder
Dimension: 768
Heads: 12

Best for:

• Better accuracy needs
• Professional work
• When speed isn’t critical

Trade-offs:

• Slower than Tiny/Base
• Higher memory requirements

---

Medium

Name: Whisper Medium
Parameters: 769 million
File Size: 1.5 GB
Memory Usage: ~5 GB
Layers: 24 encoder, 24 decoder
Dimension: 1024
Heads: 16

Best for:

• Professional transcription
• Challenging audio conditions
• Accented speech

Trade-offs:

• Significant memory usage
• Slower processing
• Requires 8GB+ RAM

---

Large-v3

Name: Whisper Large v3
Parameters: 1550 million
File Size: 3 GB
Memory Usage: ~10 GB
Layers: 32 encoder, 32 decoder
Dimension: 1280
Heads: 20

Best for:

• Maximum accuracy
• Difficult audio
• Professional production

Trade-offs:

• Very high memory usage
• Slowest processing
• Requires 16GB+ RAM

---

English-Only Models

Models ending in .en are optimized for English only:

Model	Multilingual	English-Only
Tiny	tiny	tiny.en
Base	base	base.en
Small	small	small.en
Medium	medium	medium.en
Large	large-v3	(No .en variant)

Advantages of .en Models

• Faster processing - No language detection
• Slightly better accuracy - Optimized for English
• Lower resource usage - Smaller effective vocabulary

Tip

If you only transcribe English, use .en models for best performance.

---

Parakeet Models

NVIDIA’s Parakeet models via FluidAudio framework.

Parakeet RNNT

Name: Parakeet RNNT
Architecture: RNN-Transducer
Focus: Real-time English ASR
Latency: Very low
Streaming: Yes

Best for:

• Real-time transcription
• Live dictation
• English content

Parakeet vs Whisper

Aspect	Whisper	Parakeet
Languages	99+	English focus
Accuracy	Higher	Good
Latency	Higher	Lower
Streaming	Limited	Native
Memory	Higher	Lower

---

Model Performance by Hardware

Apple Silicon

M1

Model	Recommended	Performance
Tiny	✓ Excellent	~30x real-time
Base	✓ Excellent	~15x real-time
Small	✓ Good	~5x real-time
Medium	⚠️ Usable	~2x real-time
Large	⚠️ Slow	~0.5x real-time

M1 Pro/Max

Model	Recommended	Performance
Tiny	✓ Excellent	~40x real-time
Base	✓ Excellent	~20x real-time
Small	✓ Excellent	~8x real-time
Medium	✓ Good	~3x real-time
Large	✓ Usable	~1x real-time

M2/M3

Model	Recommended	Performance
Tiny	✓ Excellent	~45x real-time
Base	✓ Excellent	~22x real-time
Small	✓ Excellent	~9x real-time
Medium	✓ Good	~4x real-time
Large	✓ Usable	~1.2x real-time

M2/M3 Pro/Max

Model	Recommended	Performance
Tiny	✓ Excellent	~50x real-time
Base	✓ Excellent	~25x real-time
Small	✓ Excellent	~12x real-time
Medium	✓ Excellent	~5x real-time
Large	✓ Good	~2x real-time

Intel Macs

Model	8GB RAM	16GB RAM	32GB+ RAM
Tiny	✓ Good	✓ Good	✓ Good
Base	✓ Usable	✓ Good	✓ Good
Small	⚠️ Slow	✓ Usable	✓ Good
Medium	❌ Not recommended	⚠️ Slow	✓ Usable
Large	❌ Not recommended	❌ Not recommended	⚠️ Slow

---

Language Support

Tier 1: Best Support (WER < 5%)

English, Spanish, French, German, Italian, Portuguese, Dutch, Russian, Japanese, Chinese, Korean

Tier 2: Good Support (WER 5-15%)

Arabic, Czech, Danish, Finnish, Greek, Hebrew, Hindi, Hungarian, Indonesian, Norwegian, Polish, Romanian, Swedish, Thai, Turkish, Ukrainian, Vietnamese

Tier 3: Basic Support (WER > 15%)

All other 70+ languages supported by Whisper

Note

WER (Word Error Rate) varies by accent, audio quality, and speaking style.

---

Audio Specifications

Input Requirements

Specification	Value
Sample Rate	16000 Hz
Bit Depth	16-bit
Channels	Mono
Format	PCM

FloWords automatically converts audio to these specifications.

Supported Input Formats

Format	Extension	Notes
WAV	.wav	Native support
MP3	.mp3	Converted to WAV
M4A	.m4a	Converted to WAV
AAC	.aac	Converted to WAV
FLAC	.flac	Converted to WAV
AIFF	.aiff	Converted to WAV
CAF	.caf	Converted to WAV
MP4	.mp4	Audio extracted
MOV	.mov	Audio extracted

---

Model Selection Guide

By Use Case

Use Case	Recommended Model
Quick notes	Tiny
Daily use	Base
Documents	Small
Professional	Medium
Maximum accuracy	Large-v3

By RAM Available

RAM	Maximum Model
4 GB	Base
8 GB	Small
16 GB	Medium
32 GB+	Large-v3

By Content Type

Content	Recommended
Clear speech	Any
Background noise	Medium+
Technical terms	Medium+ with dictionary
Multiple speakers	Medium+
Accented speech	Medium+

---

Advanced Configuration

Model Parameters

Parameter	Range	Default	Effect
beam_size	1-10	5	Accuracy vs speed
best_of	1-5	1	Candidates considered
temperature	0.0-1.0	0.0	Prediction randomness
patience	0.0-2.0	1.0	Early stopping
length_penalty	0.0-2.0	1.0	Length bias

When to Adjust

Goal	Adjustment
Faster	Lower beam_size
More accurate	Higher beam_size, best_of
More variety	Higher temperature
Shorter outputs	Lower length_penalty

---

Model Updates

Checking for Updates

FloWords checks for model updates automatically. To manually check:

1. Open Settings > Model
2. Click Check for Updates

Update Process

Model updates may include:

• Accuracy improvements
• New language support
• Performance optimizations
• Bug fixes

---

Next Steps

• Download Models to get started
• Configure Settings for your needs
• Review Best Practices for optimal use