Measure how long time it takes to synthesize utterances using the TTS engines currently available on the development server. This should give an idea of what will be needed from the servers when scaling up to more users.
Description
Details
Subject | Repo | Branch | Lines +/- | |
---|---|---|---|---|
Speechoid/Wikispeech benchmark maintenance script | mediawiki/extensions/Wikispeech | master | +480 -0 |
Status | Subtype | Assigned | Task | ||
---|---|---|---|---|---|
Open | None | T264842 Deploy Wikispeech in production | |||
Open | None | T180015 ☂ Deploy Wikispeech on beta cluster | |||
Open | None | T264748 ☂ Speechoid WMF deployment | |||
Resolved | • kalle | T264403 ☂ Benchmark maintenance script | |||
Resolved | • kalle | T247282 Estimate synthesis time |
Event Timeline
Results from manual testing:
Synthesizing about 120 words, one semi large paragraph of text, takes about 7 seconds on my laptop. It consumes 3 CPUs (out of 6 CPUs according to /proc/cpuinfo) for most of the time, indicating it somehow does this multi threaded, even though there is nothing regarding this in the Mary documentation. It might just be the JVM that automatically handle this. Setting up rate limit using HAProxy correct (see T247390) might make this speedier under heavy load at 2 connections at the time compared to, lets say, 10 connections that have to share all CPUs running at 100% each.
Change 631719 had a related patch set uploaded (by Karl Wettin (WMSE); owner: Karl Wettin (WMSE)):
[mediawiki/extensions/Wikispeech@master] Speechoid/Wikispeech benchmark maintenance script
Change 631719 merged by jenkins-bot:
[mediawiki/extensions/Wikispeech@master] Speechoid/Wikispeech benchmark maintenance script