doc.scripts.Narrator-multiplettsslaves.html Maven / Gradle / Ivy
Running speechgen2 using multiple TTS slaves
Running speechgen2 using multiple TTS slaves
Configuring multiple TTS slaves on a local system
Running multiple concurrent TTS slaves on a local system only requires a bit
of tweaking in the se_tpb_speechgen2/tts/ttsbuilder.xml config
file. Simply make sure there are several instances of the the
<param name="command" value="..."/> element within the
<tts> you which to run concurrently.
The TTS slave instances will receive one sentence to process at a time, and as
a slave instance finishes processing a sentence it will receive the next request.
Configuring distributed TTS slaves
This section describes the solution in use at TPB for using distributed TTS slaves.
Other solutions are possible, such as running a server on each of the slaves processing
the requests from Daisy Pipeline.
Protocol for communicating with the slaves
All communication with the slaves is performed using STDIN and STDOUT.
Each slave reads a single sentence on STDIN and generates the corresponding wav audio.
It then writes the length (in bytes) of the wav audio to STDOUT followed by the binary
audio data itself. The process is repeated until it reads an empty line which causes
the slave to exit.
The TTS adapter (the master) is responsible for sending the text sentences and receiving
the wav audio data from the slaves.
Using SSH for passwordless login
In order to run the TTS slave programs on remote systems, the communication is tunneled through SSH.
The SSH login must be passwordless in order to be able to run without human intervention.
A keypair needs to be generated without a passphrase on the private key. The public key is
installed in ~/.ssh/authorized_keys on the slave machine(s).
A bash script starts an SSH agent and adds the private key
before pipeline itself is executed. This way, Daisy Pipeline will be able to execute
commands on the remote machine(s) without providing a password:
...
eval `ssh-agent -s`
ssh-add /path/to/private/key
java org.daisy.pipeline.ui.CommandLineUI [parameters...]
kill -KILL $SSH_AGENT_PID
...
Some SSH installations are picky about the file permissions on the private key. Make sure
the owner of the file is the only one having read priveliges to it.
Sample ttsbuilder.xml config
In the ttsbuilder.xml config file, the command parameters executes the TTS slave programs on
the remote machines using SSH:
<tts>
<param name="command" value="ssh [email protected] tclsh current/narraFil2.tcl"/>
<param name="command" value="ssh [email protected] tclsh current/narraFil2.tcl"/>
<param name="command" value="ssh [email protected] tclsh current/narraFil2.tcl"/>
<param name="command" value="ssh [email protected] tclsh current/narraFil2.tcl"/>
<param name="class" value="se_tpb_speechgen2.tts.adapters.FilibusterTTS"/>
<param name="timeout" value="1200000"/>
...
</tts>
In this example, four TTS instances are run distributed over three systems with the third
system running two TTS instances and the other systems running one TTS instance each.
The optimal number of instances to run on each system depends on the amount of system
resources needed by each instance. TPB are currently running 15 TTS instances distributed
over three systems.