ChatGPT as a Phone assistant

Example : Recording: “For credit cards please say CREDIT CARDS, to speak with a representative please say: REPRESENTATIVE ” 

That how would look like in a simple STT use. With GPT is even better, however if someone wants to implement his own solution can use the available tools in machine learning to train his own small language model on a specific use-case to convert long phrases to paths to route the calls, having said that still is better to have a more human-like conversation by using GPT by openai as is able to solve complex problems by it self, suggest solutions to clients problems.. or even look up with API a database see which product the client purchased and provide information and how-to-use based on the manual of that specific product.

Here is an example of something like that with prompt engineering a decent prompt look like this:

You are now a call assistant who transfer calls out of you to the right departments for  EletroShop company, datetime is: 4/6/2023 17:21

You reply to a script strictly in the form of EXT:num or null:REPLY=<your reply> one cmd at the time,

EXT variable can hold extension departments:

If they need something about sales: 2001, for support: 2002, for accounting: 2003. 

If they ask for persons: John = 3000, Alex = 3030, Evelina = 3032.

CLIENT variable is the input of the client, you don’t use it. 

example: CLIENT: Hi,

You: EXT:NULL:Hi, how can i help?

CLIENT:  tv dont work

You: EXT:2002:Sorry for that, connecting you to support to handle it. please wait.

If client wants to locate order output: “EXT:NULL:LOCATE=<order_number>”

If you get an input like: ORDER-LOCATED:<text>, outptu: EXT:NULL:REPLY=Your order status is <text>

End of instructions. New session:

CLIENT: Hi.. hello

Tested with ChatGPT the results are:

And locating an order with API from the ecommerce will look like:

As you see in the above example, a client wants to find out why haven’t received his order, chatted with the language model, the language model used the “script” to initiate an API call to our backend ecommerce, then it used the result of information it got back, explained to the client successfully where his package is and plus it is aware about ACS Greek courier so it offered further help to provide contact details of the courier company. In a nutshell. just saved a couple of worthy minutes or more from the work load, other than checking the order status one can use that approach to even feed manuals for troubleshooting.

#!/usr/bin/env php

require(‘phpagi.php’);

require ‘vendor/autoload.php’;

use GoogleCloudSpeechV1SpeechClient;

use GoogleCloudSpeechV1RecognitionAudio;

use GoogleCloudSpeechV1RecognitionConfig;

use GoogleCloudSpeechV1RecognitionConfigAudioEncoding;

error_reporting(E_ALL);

$agi = new AGI();

$agi->answer();

$callerId = $agi->get_variable(“CALLERID(num)”);

$prompt = “

You are on the phone, (reply on callers language fluently) datetime is: “. date(“F j, Y, g:i a”) .” phoneNum: “.$callerId.” your name is Jarvis, n

CALLER SAID: “;

 $agi->stream_file(‘/var/lib/asterisk/agi-bin/intro’);

 sleep(1);

 $forward = “NONE”;

 while ($forward == “NONE”) {

$result = $agi->record_file(‘/tmp/’ . $agi->request[‘agi_uniqueid’], ‘wav’, ‘#’,  ‘9000’, ‘0’, 1, ‘1’);

$transcription = get_transcript(‘/tmp/’ . $agi->request[‘agi_uniqueid’] . ‘.wav’);

$text = get_ai_response($transcription);

$spoken = explode(“REPLY:”,$text);

$forward =  explode(“CMD:”,$text);

$forward =  explode(” “,$forward[1]);

$forward =  $forward[0];

 $wordsSpoken = get_sound_from_text($spoken[1]);

 $agi->stream_file($wordsSpoken,“#”);

 unlink(‘/tmp/’ . $wordsSpoken);

 unlink(‘/tmp/’ . $agi->request[‘agi_uniqueid’]);

 sleep(1);

 //

 }

 if (is_numeric($forward)) {

//here we check if CMD:num num is an extension (digits)

  $agi->exec_goto(‘from-internal’, $forward,1);

  die();

 }

 $agi->stream_file(‘/var/lib/asterisk/agi-bin/antio’);

 $agi->hangup();

  function get_transcript($audioFilePath) {

    // replace with the path to your private key JSON file

    $keyFilePath = ‘/var/lib/asterisk/agi-bin/key.json’;

    putenv(‘GOOGLE_APPLICATION_CREDENTIALS=’ . $keyFilePath);

    // replace with the path to your audio file

    $client = new SpeechClient();

    $audio = (new RecognitionAudio())

        ->setContent(file_get_contents($audioFilePath));

    $config = new RecognitionConfig([

        ‘encoding’ => AudioEncoding::ENCODING_UNSPECIFIED,

        ‘sample_rate_hertz’ => 8000,

        ‘language_code’ => ‘el-GR’

    ]);

    $response = $client->recognize($config, $audio);

    foreach ($response->getResults() as $result) {

        $alternatives = $result->getAlternatives();

        $mostLikely = $alternatives[0];

        $transcription = $mostLikely->getTranscript();

        @unlink(‘/var/lib/asterisk/agi-bin/said.txt’);

     file_put_contents(‘/var/lib/asterisk/agi-bin/said.txt’, $transcription, FILE_APPEND | LOCK_EX);

     $client->close();

     return  $transcription;

    }

    }

    function get_sound_from_text($text) {

      global $agi;

      $tmpName = $agi->request[‘agi_uniqueid’] . “-“ . rand(10,10000) . “.mp3”;

      $subscriptionKey = “your-KEY”;

      $ssml = “<speak version=’1.0′ xml:lang=’el-GR’> <voice xml:lang=’el-GR’ xml:gender=’Male’ name=’el-GR-NestorasNeural’> “. $text . ” </voice> </speak>”;

      $curl = curl_init();

      curl_setopt_array($curl, array(

        CURLOPT_URL => “https://westeurope.tts.speech.microsoft.com/cognitiveservices/v1”,

        CURLOPT_RETURNTRANSFER => true,

        CURLOPT_ENCODING => “”,

        CURLOPT_MAXREDIRS => 10,

        CURLOPT_TIMEOUT => 0,

        CURLOPT_FOLLOWLOCATION => true,

        CURLOPT_HTTP_VERSION => CURL_HTTP_VERSION_1_1,

        CURLOPT_CUSTOMREQUEST => “POST”,

        CURLOPT_POSTFIELDS => $ssml,

        CURLOPT_HTTPHEADER => array(

          “Ocp-Apim-Subscription-Key: $subscriptionKey“,

          “Content-Type: application/ssml+xml”,

          “X-Microsoft-OutputFormat: audio-16khz-128kbitrate-mono-mp3”,

          “User-Agent: curl”

        ),

      ));

      $response = curl_exec($curl);

      $err = curl_error($curl);

      curl_close($curl);

      if ($err) {

        file_put_contents(“/tmp/error_stt.txt”, “cURL Error #:” . $err);

      } else {

        file_put_contents(“/tmp/”.$tmpName, $response);

      }

     $newfln =  str_replace(“.mp3”,“.wav”, $tmpName);

         exec(“sox /tmp/$tmpName -c 1 -r 8000 /tmp/$newfln“);

         unlink(“/tmp/$tmpName“);

      return “/tmp/” . str_replace(“.wav”,“”, $newfln);

     }

 

function get_ai_response($text) {

  file_put_contents(“/tmp/text.txt”, $text);

      $curl = curl_init();

global $prompt;

curl_setopt_array($curl, array(

  CURLOPT_URL => “https://api.openai.com/v1/completions”,

  CURLOPT_RETURNTRANSFER => true,

  CURLOPT_ENCODING => “”,

  CURLOPT_MAXREDIRS => 10,

  CURLOPT_TIMEOUT => 0,

  CURLOPT_FOLLOWLOCATION => true,

  CURLOPT_HTTP_VERSION => CURL_HTTP_VERSION_1_1,

  CURLOPT_CUSTOMREQUEST => “POST”,

  CURLOPT_POSTFIELDS => json_encode(array(

    “model” => “text-davinci-003”,

    “prompt” => $prompt . $text,

    “temperature” => 0,

    “max_tokens” => 300,

    “top_p” => 0,

    “frequency_penalty” => 0,

    “presence_penalty” => 0

  )),

  CURLOPT_HTTPHEADER => array(

    “Content-Type: application/json”,

    “Authorization: Bearer  your-KEY”

  ),

));

$response = curl_exec($curl);

curl_close($curl);

$response = json_decode($response,true);

 file_put_contents(“/tmp/ai.txt”, $response[“choices”][0][“text”]);

return  $response[“choices”][0][“text”];

     }