Automatic Speech Recognition (ASR)

1# Automatic Speech Recognition (ASR)
2 
3## Overview
4 
5After speaker diarization, you need to transcribe each speech segment to text. Whisper is the current state-of-the-art for ASR, with multiple model sizes offering different trade-offs between accuracy and speed.
6 
7## When to Use
8 
9- After speaker diarization is complete
10- Need to generate speaker-labeled transcripts
11- Creating subtitles from audio segments
12- Converting speech segments to text
13 
14## Whisper Model Selection
15 
16### Model Size Comparison
17 
18| Model | Size | Speed | Accuracy | Best For |
19|-------|------|-------|----------|----------|
20| tiny | 39M | Fastest | Lowest | Quick testing, low accuracy needs |
21| base | 74M | Fast | Low | Fast processing with moderate accuracy |
22| small | 244M | Medium | Good | **Recommended balance** |
23| medium | 769M | Slow | Very Good | High accuracy needs |
24| large-v3 | 1550M | Slowest | **Best** | **Maximum accuracy** |
25 
26### Recommended: Use `small` or `large-v3`
27 
28**For best accuracy (recommended for this task):**
29```python
30import whisper
31 
32model = whisper.load_model("large-v3")  # Best accuracy
33result = model.transcribe(audio_path)
34```
35 
36**For balanced performance:**
37```python
38import whisper
39 
40model = whisper.load_model("small")  # Good balance
41result = model.transcribe(audio_path)
42```
43 
44## Faster-Whisper (Optimized Alternative)
45 
46For faster processing with similar accuracy, use `faster-whisper`:
47 
48```python
49from faster_whisper import WhisperModel
50 
51# Use small model with CPU int8 quantization
52…