TwiML™ Voice: <Gather>

You can use TwiML's <Gather> verb to collect digits or transcribe speech during a call.

The following example shows the most basic use of <Gather> TwiML:

1
<?xml version="1.0" encoding="UTF-8"?>
2
<Response>
3
    <Gather/>
4
</Response>

You can always send Twilio plain TwiML, or leverage the helper libraries to add TwiML to your web applications:

1
const VoiceResponse = require('twilio').twiml.VoiceResponse;
2

3

4
const response = new VoiceResponse();
5
response.gather();
6

7
console.log(response.toString());

1
<?xml version="1.0" encoding="UTF-8"?>
2
<!-- page located at http://example.com/simple_gather.xml -->
3
<Response>
4
    <Gather/>
5
</Response>

When Twilio executes this TwiML, it will pause for up to five seconds to wait for the caller to enter digits on their keypad. A few things might happen next:

• The caller enters digits followed by a # symbol or 5 seconds of silence. Twilio then sends the user's input as a parameter in a POST request to the URL hosting the <Gather> TwiML (if no action attribute is provided) or to the action URL.
• The caller doesn't enter any digits and 5 seconds pass. Twilio will move on to the next TwiML verb in the document - since there are no more verbs here, Twilio will end the call.

To enable V2 of Google Cloud's STT as the default on your Twilio account and override your existing <Gather> Speech configuration, do the following:

1. Navigate to Voice Settings > General in the Twilio Console.
2. Select Enabled.
3. Click Save.

Once enabled, Twilio will map your existing <Gather> attributes to use Google STT V2 as the underlying speech-to-text provider. To explicitly specify a supported Google STT V2 model, see the speechModel attribute.

<Gather> supports the following attributes that change its behavior. Review the table for updates related to Google STT V2 and the ability to specify a provider (googlev2 or deepgram) on the speechModel attribute.

Use one or more of these attributes in a <Gather> verb like so:

1
<?xml version="1.0" encoding="UTF-8"?>
2
<Response>
3
    <Gather input="speech dtmf" timeout="3" numDigits="1">
4
        <Say>Please press 1 or say sales for sales.</Say>
5
    </Gather>
6
</Response>

The action attribute takes an absolute or relative URL as a value. When the caller finishes entering digits (or the timeout is reached), Twilio will make an HTTP request to this URL. That request will include the user's data and Twilio's standard request parameters.

If you do not provide an action parameter, Twilio will POST to the URL that houses the active TwiML document.

Twilio may send some extra parameters with its request after the <Gather> ends:

If you gather digits from the caller, Twilio will include the Digits parameter containing the numbers your caller entered.

If you specify speech as an input with input="speech", Twilio will include SpeechResult and Confidence:

• SpeechResult contains the transcribed result of your caller's speech.
• Confidence contains a confidence score between 0.0 and 1.0. A higher confidence score means a better likelihood that the transcription is accurate.

Note: Your code should not expect confidence as a required field as it is not guaranteed to be accurate, or even set, in any of the results.

After <Gather> ends and Twilio sends its request to your action URL, the current call will continue using the TwiML you send back from that URL. Because of this, any TwiML verbs that occur after your <Gather> are unreachable.

However, if the caller did not enter any digits or speech, call flow would continue in the original TwiML document.

Imagine you have the following TwiML hosted at http://example.com/complex_gather.xml:

1
<?xml version="1.0" encoding="UTF-8"?>
2
<Response>
3
    <Gather>
4
        <Say>
5
            Please enter your account number,
6
            followed by the pound sign
7
        </Say>
8
    </Gather>
9
    <Say>We didn't receive any input. Goodbye!</Say>
10
</Response>

Scenario 1: If the caller:

• does not press the keypad or say anything for five seconds, or
• enters '#' (the default finishOnKey value) before entering any other digits

then they will hear, "We didn't receive any input. Goodbye!"

Scenario 2: If the caller:

• enters a digit while the call is speaking "Please enter your account number..."

then the <Say> verb will stop speaking and wait for the user's action.

Scenario 3: If the caller:

• enters 12345 and then presses #, or
• allows 5 seconds to pass

then Twilio will submit the digits and request parameters to the URL hosting this TwiML (http://example.com/complex_gather.xml). Twilio will fetch this same TwiML again and execute it, getting the caller stuck in this <Gather> loop.

To avoid this behavior, it's best practice to point your action URL to a new URL that hosts some other TwiML for handling the duration of the call.

The following code sample is almost identical to the TwiML above, but we've added the action and method attributes:

1
const VoiceResponse = require('twilio').twiml.VoiceResponse;
2

3

4
const response = new VoiceResponse();
5
const gather = response.gather({
6
    action: '/process_gather.php',
7
    method: 'GET'
8
});
9
gather.say('Please enter your account number,\nfollowed by the pound sign');
10
response.say('We didn\'t receive any input. Goodbye!');
11

12
console.log(response.toString());

1
<?xml version="1.0" encoding="UTF-8"?>
2
<!-- page located at http://example.com/complex_gather.xml -->
3
<Response>
4
    <Gather action="/process_gather.php" method="GET">
5
        <Say>
6
            Please enter your account number,
7
            followed by the pound sign
8
        </Say>
9
    </Gather>
10
    <Say>We didn't receive any input. Goodbye!</Say>
11
</Response>

Now when the caller enters their input, Twilio will submit the digits and request parameters to the process_gather.php URL.

If we wanted to read back this input to the caller, our code hosted at /process_gather.php might look like:

1
<?php
2
// page located at http://yourserver/process_gather.php
3
echo "<?xml version=\"1.0\" encoding=\"UTF-8\"?>\n";
4
echo "<Response><Say>You entered " . $_REQUEST['Digits'] . "</Say></Response>";
5
?>

Back to attributes list

finishOnKey lets you set a value that your caller can press to submit their digits.

For example, if you set finishOnKey to # and your caller enters 1234#, Twilio will immediately stop waiting for more input after they press #.

Twilio will then submit Digits=1234 to your action URL (note that the # is not included).

Allowed values for this attribute are:

• # (this is the default value)
• *
• Single digits 0-9
• An empty string ('')

If you use an empty string, <Gather> will capture all user input and no key will end the <Gather>. In this case, Twilio submits the user's digits to the action URL only after the timeout is reached.

Back to attributes list

1
<?xml version="1.0" encoding="UTF-8"?>
2
<Response>
3
    <Gather input="speech dtmf" finishOnKey="#" timeout="5">
4
        <Say>
5
            Please say something or press * to access the main menu
6
        </Say>
7
    </Gather>
8
    <Say>We didn't receive any input. Goodbye!</Say>
9
</Response>

You can improve Twilio's recognition of the words or phrases you expect from your callers by adding hints to your <Gather>.

The hints attribute contains a list of words or phrases that Twilio should expect during recognition. Supported class tokens or keywords will vary according to your provider and version:

• Google STT V1 Supported class tokens
• Google STT V2 Supported class tokens
• Deepgram keywords

You may provide up to 500 words or phrases in this list, separating each entry with a comma. Your hints may be up to 100 characters each, and you should separate each word in a phrase with a space, e.g.:

hints="this is a phrase I expect to hear, keyword, product name, name"

We have implemented Google's V1 and V2class tokens list to improve recognition. You can pass a class token directly in the hints.

hints="$OOV_CLASS_ALPHANUMERIC_SEQUENCE"

Hints are passed as keywords to Deepgram.

Back to attributes list

Specify which inputs (DTMF or speech) Twilio should accept with the input attribute.

The default input for <Gather> is dtmf. You can set input to dtmf, speech, or dtmf speech.

If you set dtmf speech for your input, the first detected input (speech or dtmf) will take precedence. If speech is detected first, finishOnKey will be ignored.

The following example shows a <Gather> that specifies speech input from the user. When this TwiML executes, the caller will hear the <Say> prompt. Twilio will then collect speech input for up to 60 seconds.

Once the caller stops speaking for five seconds, Twilio posts their transcribed speech to your action URL.

1
const VoiceResponse = require('twilio').twiml.VoiceResponse;
2

3

4
const response = new VoiceResponse();
5
const gather = response.gather({
6
    input: 'speech',
7
    action: '/completed'
8
});
9
gather.say('Welcome to Twilio, please tell us why you\'re calling');
10

11
console.log(response.toString());

1
<?xml version="1.0" encoding="UTF-8"?>
2
<!-- page located at http://example.com/simple_gather.xml -->
3
<Response>
4
    <Gather input="speech" action="/completed">
5
           <Say>Welcome to Twilio, please tell us why you're calling</Say>
6
        </Gather>
7
</Response>

Back to attributes list

The language attribute specifies the language Twilio should recognize from your caller.

This value defaults to en-US, but you can set it to any supported language of your provider.

You can set your language to any of our supported languages. See the full list.

Google STT V2 and Deepgram models map to specific languages. Please use the appropriate language supported by the selected speechModel. The language attribute is passed through to the provider unchanged.

Back to attributes list

The method you set on <Gather> tells Twilio whether to request your action URL via HTTP GET or POST.

POST is <Gather>'s default method.

Back to attributes list

You can set the number of digits you expect from your caller by including numDigits in <Gather>.

For example, you might wish to set numDigits="5" when asking your caller to enter their 5-digit zip code. Once the caller enters the final digit of 94117, Twilio will immediately submit the data to your action URL.

Back to attributes list

If you provide a partialResultCallback URL, Twilio will make requests to this URL in real-time as it recognizes speech. These requests will contain a parameter labeled UnstableSpeechResult which contains partial transcriptions. These transcriptions may change as the speech recognition progresses.

Back to attributes list

The profanityFilter specifies if Twilio should filter profanities out of your speech transcription. This attribute defaults to true, which replaces all but the initial character in each filtered profane word with asterisks, e.g., 'f***.'

If you set this attribute to false, Twilio will no longer filter any profanities in your transcriptions.

Back to attributes list

When collecting speech from your caller, speechTimeout sets the limit (in seconds) that Twilio will wait after a pause in speech before it stops its recognition. After this timeout is reached, Twilio will post the speechResult to your action URL.

If you set speechTimeout to auto, Twilio will stop speech recognition when there is a pause in speech and return the results immediately.

Back to attributes list

timeout allows you to set the limit (in seconds) that Twilio will wait for the caller to press another digit or say another word before it sends data to your action URL.

For example, if timeout is 3, Twilio wait three seconds for the caller to press another key or say another word before submitting their data.

Twilio will wait until all nested verbs execute before it begins the timeout period.

The default timeout value is 5.

Back to attributes list

speechModel allows you to select a specific model that is best suited for your use case to improve the accuracy of speech to text.

Starting October 1, 2024, you can specify a provider along with a supported model from Google Speech-to-Text (STT) V2 (googlev2) or Deepgram (deepgram). The following Google STT V2 models are currently not supported:

• chirp
• chirp_telephony
• chirp_2

The speechModel has to be specified in the format {provider_model}, where provider is either googlev2 or deepgram and the model is a supported model from Google Speech-to-Text (STT) V2 or Deepgram. See below for examples.

See Google STT V2's mapping of BCP-47 language and models for more details.

1
<Gather input="speech" speechModel="googlev2_telephony">
2
    <Say>Please tell us why you're calling.</Say>
3
</Gather>

See Deepgram's language and model mapping for more details.

1
<Gather input="speech" speechModel="deepgram_nova-2">
2
    <Say>Please tell us why you're calling.</Say>
3
</Gather>

For users on Google STT V1, the attribute supports default, numbers_and_commands, phone_call,experimental_conversations, and experimental_utterances.

numbers_and_commands and phone_call are best suited for the use cases where you'd expect to receive short queries such as voice commands or voice search. phone_call is best for audio that originated from a PSTN phone call (typically an 8khz sample rate).

1
<Gather input="speech" enhanced="true"speechModel="phone_call">
2
    <Say>Please tell us why you're calling.</Say>
3
</Gather>

Experimental models are designed to give access to the latest speech technology and machine learning research, and can provide higher accuracy for speech recognition over other available models. However, some features that are supported by other available models are not yet supported by the experimental models such as confidence scores.

The experimental_utterances model is for short utterances that are a few seconds in length and is useful for trying to capture commands or other single shot directed speech use cases; think "press 0 or say 'support' to speak with an agent." The experimental_conversations model supports spontaneous speech and conversations; think "tell us why you're calling today."

Both experimental_conversations and experimental_utterances values for speechModel currently only supports a set of languages.

Please explore all options to see which works best for your use case.

Back to attributes list

The enhanced attribute instructs <Gather> to use a premium speech model that will improve the accuracy of transcription results. The premium speech model is only supported with the phone_call model in Google STT V1. It costs 50% more at $0.03 per 15s of <Gather> than the standard phone_call model. The premium phone_call model was built using thousands of hours of training data. It ensures 54% fewer errors when transcribing phone conversations when compared to the basic phone_call model.

The following TwiML instructs <Gather> to use premium phone_call model:

1
<Gather input="speech" enhanced="true"speechModel="phone_call">
2
    <Say>Please tell us why you're calling.</Say>
3
</Gather>
4

1
<Gather input="speech" enhanced="true" speechModel="numbers_and_commands">
2
    <Say>Please tell us why you're calling.</Say>
3
</Gather>
4

Read the Language appendix to see if the enhanced model is available for your language.

Back to attributes list

actionOnEmptyResult allows you to force <Gather> to send a webhook to the action URL even when there is no DTMF input. By default, if <Gather> times out while waiting for DTMF input, it will continue on to the next TwiML instruction.

For example, in the following TwiML when <Gather> times out, <Say> instruction is executed.

1
<?xml version="1.0" encoding="UTF-8"?>
2
<Response>
3
    <Gather>
4
        <Say>
5
            Please enter your account number,
6
            followed by the pound sign
7
        </Say>
8
    </Gather>
9
    <Say>We didn't receive any input. Goodbye!</Say>
10
</Response>

To always force <Gather> to send a webhook to the action URL, use the following TwiML:

1
<?xml version="1.0" encoding="UTF-8"?>
2
<Response>
3
    <Gather actionOnEmptyResult="true" action="/gather-action">
4
        <Say>
5
            Please enter your account number,
6
            followed by the pound sign
7
        </Say>
8
    </Gather>
9
</Response>

You can nest the following verbs within <Gather>:

• <Pause>
• <Play>
• <Say>

The following example shows a <Gather> with a nested <Say>. This will read some text to the caller, and allows the caller to enter input at any time while that text is read to them:

1
const VoiceResponse = require('twilio').twiml.VoiceResponse;
2

3

4
const response = new VoiceResponse();
5
const gather = response.gather({
6
    input: 'speech dtmf',
7
    timeout: 3,
8
    numDigits: 1
9
});
10
gather.say('Please press 1 or say sales for sales.');
11

12
console.log(response.toString());

1
<?xml version="1.0" encoding="UTF-8"?>
2
<Response>
3
    <Gather input="speech dtmf" timeout="3" numDigits="1">
4
        <Say>Please press 1 or say sales for sales.</Say>
5
    </Gather>
6
</Response>

If you are using <Play> verbs, we recommend hosting your media in AWS S3 in us-east-1, eu-west-1, or ap-southeast-2 depending on which Twilio Region you are using. No matter where you host your media files, always ensure that you're setting appropriate Cache Control headers. Twilio uses a caching proxy in its webhook pipeline and will cache media files that have cache headers. Serving media out of Twilio's cache can take 10ms or less. Keep in mind that we run a fleet of caching proxies so it may take multiple requests before all of the proxies have a copy of your file in cache.

When a <Gather> reaches its timeout without any user input, call control will fall through to the next verb in your original TwiML document.

If you wish to have Twilio submit a request to your action URL even if <Gather> times out, include a <Redirect> after the <Gather> like this:

1
const VoiceResponse = require('twilio').twiml.VoiceResponse;
2

3

4
const response = new VoiceResponse();
5
const gather = response.gather({
6
    action: '/process_gather.php',
7
    method: 'GET'
8
});
9
gather.say('Enter something, or not');
10
response.redirect({
11
    method: 'GET'
12
}, '/process_gather.php?Digits=TIMEOUT');
13

14
console.log(response.toString());

1
<?xml version="1.0" encoding="UTF-8"?>
2
<!-- page located at http://example.com/gather_hints.xml -->
3
<Response>
4
    <Gather action="/process_gather.php" method="GET">
5
        <Say>Enter something, or not</Say>
6
    </Gather>
7
    <Redirect method="GET">
8
        /process_gather.php?Digits=TIMEOUT
9
    </Redirect>
10
</Response>

With this code, Twilio will move to the next verb in the document (<Redirect>) when <Gather> times out. In our example, we instruct Twilio to make a new GET request to /process_gather.php?Digits=TIMEOUT

A few common problems users face when working with <Gather>:

Problem: <Gather> doesn't receive caller input when the caller is using a VoIP phone.

Solution: Some VoIP phones have trouble sending DTMF digits. This is usually because these phones use compressed bandwidth-conserving audio protocols that interfere with the transmission of the digit's signal. Consult your phone's documentation on DTMF problems.

Problem: Twilio does not send the Digits parameter to your <Gather> URL.

Solution: Check to ensure your application is not responding to the action URL with an HTTP 3xx redirect. Twilio will follow this redirect, but won't resend the Digits parameter.
If you encounter other issues with <Gather>, reach out to our support team for assistance.

Click here to download a .csv file of all available languages.

• Eleven Best Practices, Tips, and Tricks for using Speech Recognition and Virtual Agent Bots with Voice Calling on the Twilio CPaaS Platform