Speech To Text
How does Speech To Text work 🔗
OpenVidu provides a Speech To Text module that allows transcribing in real time the audio tracks of an OpenVidu Session.
- OpenVidu is able to deliver events to the client side with the text transcription of Streams that have audio.
- Clients are able to receive events for one or multiple Streams of an OpenVidu Session, including their own Stream.
- Events are returned in real time, following a recognizing-to-recognized strategy: when a speaker that is being transcribed is talking, events flagged as recognizing will be generated one after the other as the speaker delivers one sentence. The transcribed text of recognizing events may change from one to another, while the engine gathers information about the final sentence. When the engine considers that the speaker has completed a full sentence, it triggers a recognized event with the final result.
Speech To Text engines 🔗
Azure 🔗
See Azure web.
Microsoft provides an Azure service called Speech To Text that transcribes spoken audio to text. OpenVidu seamlessly integrates the audio streams of OpenVidu Sessions with this Azure service. The only thing needed is a key for the Cognitive Service API of Azure.
Enabling Speech To Text module 🔗
OPENVIDU_PRO_SPEECH_TO_TEXT=azure
OPENVIDU_PRO_SPEECH_TO_TEXT_AZURE_KEY=<AzureKey> ## e.g. rywfyDIAL5BM70ErU9O1XSIFzWk2QQhP
OPENVIDU_PRO_SPEECH_TO_TEXT_AZURE_REGION=<AzureRegion> ## e.g. westeurope
Available languages 🔗
There are dozens of different languages supported by Azure. You have the complete list in this link. Use the language code in the table's column Speech-to-text
when subscribing to Speech to Text events.
AWS 🔗
See AWS web.
AWS provides a service called Amazon Transcribe that transcribes spoken audio to text. OpenVidu seamlessly integrates the audio streams of OpenVidu Sessions with this AWS service. The only thing needed is a set of AWS credentials with access to Amazon Transcribe.
Enabling Speech To Text module 🔗
OPENVIDU_PRO_SPEECH_TO_TEXT=aws
OPENVIDU_PRO_AWS_ACCESS_KEY=<AWS_ACCESS_KEY_ID> ## e.g. AKIAIOSFODNN7EXAMPLE
OPENVIDU_PRO_AWS_SECRET_KEY=<AWS_SECRET_ACCESS_KEY> ## e.g. wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY
OPENVIDU_PRO_AWS_REGION=<AWS_DEFAULT_REGION> ## e.g. eu-west-1
Available languages 🔗
There are multiple languages supported by AWS. You have the complete list in this link (compatible languages are the ones with option streaming
in their Data input
table's column). Use the language code in the table's column Language Code
when subscribing to Speech to Text events.
Vosk 🔗
See Vosk web.
This is the open source alternative with no extra cost for the service: you will not have to pay an external cloud provider for the service. But you must take into account that it may affect the average accuracy of the transcription and the amount of resources consumed in your OpenVidu cluster (as the engine needs to run in your Media Nodes).
Enabling Speech To Text module 🔗
Configure the following configuration property in the .env
file at OpenVidu installation path (default to /opt/openvidu
):
OPENVIDU_PRO_SPEECH_TO_TEXT=vosk
Available languages 🔗
The default languages offered by OpenVidu are the small versions of the Vosk models for each one of the languages in the table below (all have Apache 2.0 license). Note that these models are lightweight, originally intended for low-resource devices, and transcription accuracy may be affected. You can configure your own custom models for better results.
Language | Code |
---|---|
English | en-US |
Spanish | es-ES |
French | fr-FR |
German | de-DE |
Portuguese | pt-PT |
Italian | it-IT |
Dutch | nl-NL |
Catalan | ca-ES |
Japanese | ja-JP |
Chinese | zh-CN |
Hindi | hi-IN |
Using custom languages 🔗
Vosk models are available here. The default ones offered by OpenVidu are the small versions of the models for each language, which all have Apache 2.0 license.
To use a custom language model, you need to override the Speech to Text module to use a custom one. It is very easy: you just need to create a Docker image based on the default one openvidu/speech-to-text-service-base
. The only thing to do in your custom image is copying your language model to path /app/dist/vosk-models
. Let's go step by step, considering that we want to add a new language called custom-model
:
1) Go to an empty directory and create a file called Dockerfile
with the following content:
FROM openvidu/speech-to-text-service-base:2.30.0
COPY custom-model /app/dist/vosk-models/custom-model
The Docker image openvidu/speech-to-text-service-base
does not have any default language model in it. If you want to preserve the default language models, you can use openvidu/speech-to-text-service
as base image instead.
2) Download the desired language model and unzip it.
3) Rename the unzipped folder to custom-model
. You can use any other name, but in these steps we will use this one. Just replace custom-model
with your own name in the next steps if you want.
4) Copy the custom-model
folder to the same directory where the Dockerfile
is and build the Docker image:
docker build . -t <your-registry>/<image-name>:<tag>
5) Push the image to your Docker registry.
docker push <your-registry>/<image-name>:<tag>
6) Configure the new image in the .env
configuration file of OpenVidu (by default located at /opt/openvidu/.env
). You can also configure the Docker registry credentials if needed. If you are using Dockerhub as Docker registry this is not necessary.
OPENVIDU_PRO_SPEECH_TO_TEXT_IMAGE=<your-registry>/<image-name>:<tag>
OPENVIDU_PRO_DOCKER_REGISTRIES=["serveraddress=<your-registry>,username=<your-username>,password=<your-password>"]
- When subscribing to Speech to Text events in your application's client, just configure your new language (that would be
custom-model
in the steps above) as second parameter of the method. Check out section Receiving Speech To Text events.- You can add as many custom language models as you want in the same Docker image. Just store them in path
/app/dist/vosk-models
as stated in step 1)
Managing language models in Media Nodes 🔗
Vosk language models are large files that must be loaded and unloaded from memory inside the Media Nodes. Configuration property OPENVIDU_PRO_SPEECH_TO_TEXT_VOSK_MODEL_LOAD_STRATEGY
dictates the strategy that OpenVidu will follow to load/unload language models in Media Nodes:
- If
OPENVIDU_PRO_SPEECH_TO_TEXT_VOSK_MODEL_LOAD_STRATEGY=on_demand
: OpenVidu by default will dynamically load the required language model in the Media Node when it is first needed. That is: when a user has requested a Speech to Text transcription for that specific language and there is no other previous transcriptions using that language in that Media Node. Besides this, OpenVidu will dynamically unload the language model from the Media Node when it is no longer needed. That is: when the last subscriptions to a Speech to Text transcription for a specific language in a particular Media Node is terminated. - If
OPENVIDU_PRO_SPEECH_TO_TEXT_VOSK_MODEL_LOAD_STRATEGY=manual
: OpenVidu won't load or unload language models dynamically in Media Nodes. You must call the REST API methods to manually load and unload a language model in a specific Media Node when necessary. If a user tries to subscribe to a Speech to Text transcription and the requested language model is not available in the required Media Node, an error will be returned.
Why do these two options exist? Ideally, option on_demand
is the most convenient from an administration perspective: OpenVidu just takes care of it whenever it is necessary. But the reality is somewhat more complicated: since the Vosk language models are large files, it is very inefficient to dynamically load and unload them in the Media Nodes, especially if the application's use case involves many different activations and deactivations of Speech To Text transcriptions.
On top of that, the default language models provided by OpenVidu are small in size and in general will work just fine with on_demand
option, but when using custom language models that are bigger in size, the amount of time required to load the model in memory can dramatically increase. Clients calling the Speech to Text subscription operation may have to wait up to 10 seconds for the operation to return, and that is not acceptable (timeout errors may even occur). So in this case, it is highly recommended (if not almost mandatory) to use manual
option to control the loading and unloading process of the language models in your Media Nodes. So when a user requests a Speech to Text subscription, the model is already available and ready in the required Media Node.
Check out the REST API operations to manually load and to unload a Vosk language model in a Media Node.
Receiving Speech To Text events 🔗
To receive Speech To Text events in your application's client side you just need to setup listener speechToTextMessage
in the Session
object. The listener will handle SpeechToTextEvent objects when the targetted participant speaks. You can differentiate between sentences under construction or final sentences using the event property reason
:
session.on("speechToTextMessage", event => {
if (event.reason === "recognizing") {
console.log("User " + event.connection.connectionId + " is speaking: " + event.text);
} else if (event.reason === "recognized") {
console.log("User " + event.connection.connectionId + " spoke: " + event.text);
}
});
Then you just need to subscribe to the desired Stream transcription using method Session.subscribeToSpeechToText
. Pass the desired Stream object for which you want to receive Speech To Text events, and the language:
await session.subscribeToSpeechToText(stream, "en-US");
In case you are using vosk and a custom model, you need to specify the name of the model you have added in the Docker image as the second parameter of the method. For example, if your custom model is located in /app/dist/vosk-models/custom-model
inside your custom Speech to Text Docker image, you need to pass custom-model
. For example:
await session.subscribeToSpeechToText(stream, "custom-model");
For OpenVidu WebComponent and OpenVidu Angular Components, you need to add the custom model name as a new value at:
Check out tutorial openvidu-speech-to-text to test a real sample application.
Reconnecting to Speech to Text module in the case of a crash 🔗
Speech to Text is a beta feature that could experiment unexpected crashes in rare occasions. openvidu-browser SDK provides an event to know if the service has crashed, so that the application may re-establish the transcription subscriptions once it is available again (the Speech to Text module restarts on its own in case of a crash). To do so, simply listen to the ExceptionEvent in your Session object, and filter by SPEECH_TO_TEXT_DISCONNECTED
name. See the code snippet below:
session.on("exception", async (event) => {
if (event.name === "SPEECH_TO_TEXT_DISCONNECTED") {
console.warn("Speech to Text service has disconnected. Retrying the subscription...");
var speechToTextReconnected = false;
while (!speechToTextReconnected) {
await new Promise(r => setTimeout(r, 1000)); // Waiting one second
try {
await session.subscribeToSpeechToText(stream, "en-US");
console.log("Speech to Text service has recovered");
speechToTextReconnected = true;
} catch (error) {
console.warn("Speech to Text service still unavailable. Retrying again...")
}
}
} else {
// Other types of ExceptionEvents
}
});