ASR Pipeline Overview
Automatic Speech Recognition transformă audio în text. Modern ASR folosește deep learning end-to-end, dar înțelegerea componentelor ajută la optimizare.
<200ms
Target latency
<10%
Target WER
Real-time
Streaming
Processing Pipeline
1
Audio Input
Raw audio waveform (16kHz, 16-bit)
2
Feature Extraction
Convert to MFCC or filterbank features
3
Acoustic Model
Neural network maps audio to phonemes
4
Language Model
Predict most likely word sequences
5
Decoder
Combine scores, output text
Model Architectures
CTC (Connectionist Temporal Classification)
Streaming-friendly, lower accuracy
Attention-based (Transformer)
Higher accuracy, higher latency
RNN-T (Transducer)
Best balance for streaming
Whisper-style
Highest accuracy, batch processing
Feature Extraction
Audio raw → Features numerice pe care rețeaua le poate procesa:
// MFCC Feature Extraction
function extractMFCC(audio, config) {
// 1. Pre-emphasis (boost high frequencies)
const emphasized = preEmphasis(audio, 0.97);
// 2. Frame the signal (25ms windows, 10ms hop)
const frames = frame(emphasized, 400, 160);
// 3. Apply Hamming window
const windowed = frames.map(f => hamming(f));
// 4. FFT to get spectrum
const spectra = windowed.map(f => fft(f));
// 5. Apply mel filterbank (40 filters)
const melSpectra = spectra.map(s => melFilterbank(s, 40));
// 6. Log and DCT to get MFCCs
const mfccs = melSpectra.map(m => dct(log(m)));
return mfccs; // 13-40 features per frame
}Optimizations
| Optimization | Impact | Tradeoff |
|---|---|---|
| Quantization | 2-4x faster | Minimal accuracy loss |
| Pruning | 30-50% smaller | Slight accuracy drop |
| Distillation | Smaller model, same accuracy | Training cost |
| Streaming Chunks | Lower latency | May miss context |