Automatic Speech Recognition

Automatic speech recognition converts an audio signal into text. The classical pipeline had four separable parts: a front end turning waveform into spectral frames, an acoustic model scoring frames against sound units, a pronunciation lexicon, and a language model supplying priors over word sequences — historically an n-gram, which is to say a Markov Chain. Modern systems train end to end, but the division still explains the failure modes.

Streaming decoding emits partial hypotheses under a latency budget and cannot see the rest of the utterance, so it revises earlier words as evidence arrives — the visible churn in live captions. Batch decoding has the whole recording and is reliably more accurate. The choice is a latency-accuracy trade, and it shapes anything consuming a Streaming Response.

Word error rate is the standard metric: substitutions plus deletions plus insertions, over reference words. Its blind spots matter. Every word counts equally, so dropping not costs what dropping um costs. Casing, punctuation and number formatting are usually normalized away before scoring, so Character Encoding and normalization choices can move the number more than a model upgrade would. And the reference transcript is human work, so the Ground Truth has an error rate of its own — something any Evaluation Harness for speech must price in.

Real accuracy is dominated by what the headline number omits: proper nouns and rare vocabulary, which carry low prior probability and often need explicit biasing lists; disfluency, overlapping speakers and code-switching; and acoustics. The last is a hard ceiling. A far-field microphone with aggressive gain control and noise suppression has already discarded information, and no decoder recovers what is not in the signal — which is why a close-talking microphone routinely beats a larger model, and why Hallucination on near-silent audio is a known failure of end-to-end systems.