FastCGI support for Kaldi. It allows Kaldi based speech recognition to be used though Apache or Nginx (or any other that support FastCGI) HTTP servers. It also contains simple HTML-based client, that allows testing Kaldi speech recognitionfrom a web page.

Apache 2.0

This guide will help you to download and build your own simple ASR web-service based on Kaldi ASR code.

Let's create a directory where all data will be downloaded and built.

mkdir ~/apiai
cd ~/apiai

You are free to choose any other name and path you wish to, but will have to keep in mind that your name differs from the name given in the guide.

Due to server code is based on Kaldi almost all prerequisites matches to Kaldi ones. Besides that a FastCGI library is required to communicate with HTTP server.

As a first step you have to clone Kaldi source tree available at https://github.com/kaldi-asr/kaldi:

git clone https://github.com/kaldi-asr/kaldi

This command will clone source tree to kaldi directory. To configure and build Kaldi please refer to kaldi/INSTALL file. For detailed information please look for Kaldi official instruction: http://kaldi-asr.org/doc/install.html

There are some extra libraries required. You may install them using system packet manager.

In openSuSE you may run:

$ sudo zypper install FastCGI-devel

It you have Debian or Ubuntu:

$ sudo apt-get install libfcgi-dev

Return to your working directory where you put Kaldi sources

$ cd ~/apiai

and then clone server source code

$ git clone https://github.com/api-ai/asr-server asr-server

It is recommended to checkout code to the same directory where kaldi-apiai is located to allow configure tool to detect Kaldi location automatically.

$ cd asr-server

Before running a make process you have to configure build scripts by running a special utility:

$ ./configure

It will check that all required libraries installed to your system and also will look for Kaldi libraries in ../kaldi folder. If you have Kaldi installed somewhere else you may explicitly pass the path via --kaldi-root option:

$ ./configure --kaldi-root=<path_to_kaldi>

If configuration process has finished successfully you may begin the building process by running make script:

$ make

When application build complete you need to download language specific data.

Return to your working directory where you put Kaldi sources

$ cd ~/apiai

Builded ASR application uses a Kaldi nnet3 models, which you can get by training a neural network with your personal data set or use a pretrained network provided by us. Currently it is only English model available at https://github.com/api-ai/api-ai-english-asr-model/releases/download/1.0/api.ai-kaldi-asr-model.zip.

$ wget https://github.com/api-ai/api-ai-english-asr-model/releases/download/1.0/api.ai-kaldi-asr-model.zip

Unzip the archive to asr-server directory.

$ unzip api.ai-kaldi-asr-model.zip

Set the model directory as a working dir:

$ cd api.ai-kaldi-asr-model

There are several ways available to run application. The first one is to run it as a standalone app listening on socket defined with --fcgi-socket option:

$ ../asr-server/fcgi-nnet3-decoder --fcgi-socket=:8000

This command runs application listening on any IP address and port 8000. You are also free to define a path Unix socket, or explicit IP address (in a A.B.C.D:PORT form).

As an alternative way you may use special spawn-fcgi utility:

$ spawn-fcgi -n -p 8000 -- ../asr-server/fcgi-nnet3-decoder

You may use any web-server which have FastCGI support: Apache, Nginx, Lighttpd etc.

openSuSE:

$ sudo zypper in apache2

Debian and Ubuntu:

$ sudo apt-get install apache2

Enable FastCGI proxy module with a2enmod:

$ sudo a2enmod proxy_fcgi

Then you have to add to Apache2 configuration file following line:

ProxyPass "/asr" "fcgi://localhost:8000/"

If your Apache configured to include all .conf files from /etc/apache2/conf.d folder you may create separate asr_proxy.conf file with following content:

ProxyPass "/asr" "fcgi://localhost:8000/"
Alias /asr-html/ "/home/username/apiai/asr-server/asr-html/"
<Directory "/home/username/apiai/asr-server/asr-html">
	Options Indexes MultiViews
	AllowOverride None
	Require all granted
</Directory>

Now restart Apache:

$ sudo /etc/init.d/apache2 restart

You can download latest sources from official website http://nginx.org/ and build Nginx with yourself or use your system package manager.

openSuSE:

$ sudo zypper install nginx

Debian and Ubuntu:

$ sudo apt-get install nginx

Open nginx.conf and write down the following code:

http {
	server {
		location /asr {
			fastcgi_pass 127.0.0.1:8000;
			# Disabling this option invokes immediate sending replies to client
			fastcgi_buffering off;
			# Disabling this option invokes immediate decoding incoming audio data
			fastcgi_request_buffering off;
			include      fastcgi_params;
		}

		location /asr-html {
			root /home/username/apiai/asr-server/;
			index index.html;
		}
	}
}

This will setup Nginx to pass all requests coming to url /asr directly to ASR service listening 8000 port via FastCGI gate. For detailed information please please refer to nginx documentation (e.g. https://www.nginx.com/resources/wiki/start/topics/examples/fastcgiexample/)

Server accepts raw mono 16-bits 16 KHz PCM data. You can convert your audio using any popular encoding utilities, for instance, you can use ffmpeg:

$ ffmpeg -i audio.wav -f s16le -ar 16000 -ac 1 audio.raw

There is a simple JS implementation that allows you to recognize speech using system mic. Open in your browser:

http://localhost/asr-html/

and follow the instructions on the page.

Now, let’s recognize audio.raw by calling web-service with curl utility:

$ curl -H "Content-Type: application/octet-stream" --data-binary @audio.raw http://localhost/asr

On successfull recognition the command will return something like this:

{
	"status":"ok",
	"data":[{"confidence":0.900359,"text":"HELLO WORLD"}]
}

On error the return value will be like this:

{"status":"error","data":[{"text":"Failed to decode"}]}

There are several parameters to tune up recognition process. All parameters are expected to be passed via query string as web-form fields enumeration (e.g. ?name1=value1&name2=value2).

{
	"status":"ok",
	"data":[
		{"confidence":0.900359,"text":"HELLO WORLD"},
		{"confidence":0.89012,"text":"HELLO WORD"}
	]
}

{
	"status":"ok",
	"data":[{"confidence":0.900359,"text":"HELLO WORLD"}],
	"interrupted":"endofspeech",
	"time":3800
}

		The result returned as an simple sequence of JSON documents.
		Each intermediate document have "status" field set to "intermediate",
		last one will have "status" set to "ok".

{"status":"intermediate","data":[
	{"confidence":0.908981,"text":"HELLO"}
]}
{"status":"intermediate","data":[
	{"confidence":0.903025,"text":"HELLO WORLD"}
]}
{"status":"ok","data":[
	{"confidence":0.903025,"text":"HELLO WORLD"}
]}

--ResponseBoundary
Content-Disposition: form-data; name="partial"
Content-type: application/json
{"status":"intermediate","data":[
{"confidence":0.908981,"text":"HELLO"}
]}
--ResponseBoundary
Content-Disposition: form-data; name="partial"
Content-type: application/json
{"status":"intermediate","data":[
{"confidence":0.903025,"text":"HELLO WORLD"}
]}
--ResponseBoundary
Content-Disposition: form-data; name="result"
Content-type: application/json
{"status":"ok","data":[
{"confidence":0.903025,"text":"HELLO WORLD"}
]}
--ResponseBoundary--