Module bench 

The bench’s corpus: the audio a run scores against, taken from this machine’s own material and from the slice the binary carries.

Three sources, and what each can say is different. A saved take is this machine’s own speech with the in-use model’s words beside it - latency on real material, but its text.txt is a model’s output, not truth: the consensus tier reads how far the candidate sits from the model in use, which is a signal and not an error. The read-aloud set is the one source whose words are known, because they are a passage somebody read aloud on purpose - only a clip from it can be scored for term accuracy, and everything here keeps that distinction in the types.

The public slice is the exception to “nothing is embedded”. It is a fixed LibriSpeech test-clean slice carried in the binary and installed beside the models on first bench, so a machine that never recorded anything still has known words to score against.

Re-exports

use std::path::Path; 🔒

use std::time::Duration; 🔒

use sha2::Digest as _; 🔒

use sha2::Sha256; 🔒

use crate::Config; 🔒

use crate::Error; 🔒

use crate::SAMPLE_RATE; 🔒

use crate::Stages; 🔒

Modules

assets 🔒

The public bench slice that ships in the binary: LibriSpeech test-clean utterances, their transcripts, and the attribution the licence asks for.

Structs

Clip

One clip of audio, and whatever is known about it.

ClipOutcome 🔒

One clip’s transcription as the loop needs it.

ClipRun

One clip’s outcome, as the callback sees it while a run is out.

Corpus

The audio a run scores against, and its identity.

CorpusId

What two runs are comparable by: the corpus’s own shape and bytes.

Manifest 🔒

The manifest the installed public slice carries beside its clips.

ManifestClip 🔒

One clip the public manifest names: a wav, its digest, and the words it was read from.

Metrics

What one run measured.

StageTotals

Where one run’s time went, summed over its clips.

Enums

ClipSource

Where one clip came from.

Tier

Which clips a run scores against, in the names the page uses.

Constants

KNOWN_TERMS 🔒

The words a mishearing shows up in, across the whole corpus. A term survives when it comes out whole and spelled right, case-insensitively - which is the figure a bench exists to move.

READ_ALOUD_PASSAGE

The passage the read-aloud set is recorded from: the one document whose words are known, so a run over it can be scored rather than compared. Read aloud once, kept like a model artifact, and re-used for every future candidate.

Functions

corpus

Build one tier’s corpus from the machine’s own material, and the installed public slice where the tier asks for it.

corpus_id 🔒

The identity two runs are comparable by: the clip count, the seconds of audio, and a hash over every clip’s own samples, so two corpora that differ by one take cannot share an id.

decode_reader 🔒

decode_wav 🔒

Decode one 16-bit mono wav at the model’s own rate, refusing anything else by name rather than resampling it.

edits 🔒

Word-level Levenshtein distance: the edits between two word sequences.

install_public_corpus

Install the bundled public slice into models_dir’s public-corpus directory. Idempotent: a slice this binary wrote is left alone, and a manifest it did not write is replaced by the bundled one.

public_corpus

The installed public slice: every clip the manifest names, each verified against its own digest before anything scores it.

ratio

One count’s share of another, 0.0 when there is nothing to divide by.

read_aloud

The read-aloud set: the passage, and every recording of it this machine has kept - each scored against the passage as its truth.

run

Run one corpus through one config, reporting each clip as it lands.

samples_seconds 🔒

A count of samples as seconds at the model’s own rate.

score

Term accuracy and word error rate of text against truth.

store_read_aloud

Write one recording into the read-aloud set: the passage read, the take the words were recorded in, and a manifest naming the wav’s own sha256.

store_read_aloud_at 🔒

store_read_aloud for a caller that brings its own stamp, which is how two recordings in one millisecond are testable at all.

takes

Every complete take in dir, oldest first.

terms

The passage’s own terms: every word in it the vocabulary knows, whole words only, deduplicated, in the passage’s order.

walk 🔒

Walk the corpus, handing each clip’s outcome to on_clip; a Break stops the run where it stands.

wer

Word error rate of text against truth: edits over the truth’s own words, and no error when the truth has none.

words 🔒

A string’s words: lowercased, split on anything that is not a letter or a digit, blanks dropped. The one spelling both axes compare on.