debug_timings.py

# please don't commit this to user repos, it uses unstable talon APIs - in
# particular, the audio metadata.
from talon import speech_system
import logging

# ignore all phrase recognitions that take fewer than this many ms.
IGNORE_THRESHOLD = 100

# name: (bad_threshold, worse_threshold)
# you probably want to modify these for your computer & usage patterns.
TIMINGS = {'audio': (10000, 30000),
           'compile': (500, 1000),
           'emit': (200, 400),
           'decode': (30, 60),
           'total': (1000, 2000),
           }

def format_time(name, time, lo, hi):
    unesc = "" if time < lo else "\033[0m"
    esc = ("" if time < lo else
           # yellow if above lo threshold
           "\033[33m" if time < hi else
           # red if above hi threshold
           "\033[31m")
    return esc + f"{name}={time:.3f}ms" + unesc
    
def print_timings(j):
    try:
        phrase = ' '.join(j['phrase'])
        meta = j['_metadata']
        timings = [(name, meta[name+'_ms'], lo, hi)
                   for name, (lo, hi) in TIMINGS.items()]
    except KeyError: pass
    else:
        if meta['total_ms'] < IGNORE_THRESHOLD: return
        logging.info(f'PHRASE {phrase}')
        if meta['compile_ms'] > 1: logging.info('-> dfa recompile')
        logging.info(" ".join(format_time(*x) for x in timings))

speech_system.register('phrase', print_timings)