Module bench 

The speaking bench: synthesize a corpus’s transcripts, transcribe them back through a caller-supplied recognizer, and score the round trip with its first-audio time, realtime factor and the resident set it needed.

The recognizer is a parameter because this crate is a leaf and owns no model of the other direction.

Re-exports

use std::sync::Arc; 🔒

use std::sync::atomic::AtomicBool; 🔒

use std::sync::atomic::AtomicU64; 🔒

use std::sync::atomic::Ordering; 🔒

use std::time::Duration; 🔒

use std::time::Instant; 🔒

use sysinfo::Pid; 🔒

use sysinfo::ProcessesToUpdate; 🔒

use sysinfo::System; 🔒

use crate::Engine; 🔒

use crate::Error; 🔒

use crate::SpeakRequest; 🔒

Structs

BenchClip

One clip to score: the words to say.

PeakRss 🔒

A resident-set sampler that runs beside the bench for its duration.

TtsBenchResult

What one run measured.

Functions

edits 🔒

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

ratio 🔒

Edits over the truth’s own words; a truth with no words is no error rather than a division by zero.

resident_bytes 🔒

This process’s resident set, in bytes; zero where the platform will not answer.

run

Synthesize every transcript, transcribe the audio back, and score it.

words 🔒

A string’s words: lowercased, split on anything that is not a letter or a digit, blanks dropped.