Ce Este WER
Word Error Rate măsoară câte cuvinte sunt greșite. WER = (S + I + D) / N. Mai mic = mai bun. Target pentru voice AI: sub 10%.
WER = (Substitutions + Insertions + Deletions) / Total Words × 100%
Error Types
Substitutions
Wrong word recognized
"tree" → "three"Insertions
Extra words added
"hello" → "hello there"Deletions
Words missed
"I want pizza" → "I want"Optimization Techniques
Audio Quality
- ✓ Noise suppression
- ✓ Echo cancellation
- ✓ Gain normalization
- ✓ High-quality codec
Custom Vocabulary
- ✓ Add domain terms
- ✓ Boost product names
- ✓ Industry jargon
- ✓ Proper nouns
Language Model
- ✓ Domain-specific LM
- ✓ Contextual biasing
- ✓ Phrase lists
- ✓ Grammar constraints
Acoustic Adaptation
- ✓ Fine-tune on your audio
- ✓ Accent training
- ✓ Noise condition matching
Typical WER Benchmarks
| Scenario | Expected WER |
|---|---|
| Clean audio, native speaker | 5-8% |
| Phone quality (8kHz) | 8-12% |
| Noisy environment | 10-15% |
| Accented speech | 10-20% |
| Technical terminology | 15-25% |
| With custom vocab | 5-10% |
Custom Vocabulary Example
// Deepgram custom vocabulary
const transcriptionConfig = {
model: "nova-2",
language: "ro",
keywords: [
// Product names with boost
"Kallina:2",
"ChatBot:1.5",
// Industry terms
"CRM:1.5",
"API:1.5",
// Proper nouns
"București:1.5",
"Chișinău:1.5"
],
// Phrases to boost
search: [
"verifică comanda",
"programează o întâlnire",
"vreau să vorbesc cu cineva"
]
};