Skip to main content

Using real-time streaming

AssemblyAI's Streaming Speech-to-Text (STT) service allows you to transcribe live audio streams with high accuracy and low latency. By streaming your audio data to our secure WebSocket API, you can receive transcripts back within a few hundred milliseconds, and our system continues to revise these transcripts with greater accuracy over time as more context arrives.

In this guide, you'll learn how to establish a WebSocket connection, send audio data, and receive partial and final transcription results. For more information about the expected audio format, see Audio Requirements.

Get started​

Before we begin, make sure you have an AssemblyAI account and an API key. You can sign up for a free account and get your API key from your dashboard. Please note that this feature is available for paid accounts only. If you're on the free plan, you'll need to upgrade.

The entire source code of this guide can be viewed here.

Step-by-step instructions​

  1. 1

    To use the microphone stream you need to install pyaudio. Mac and Linux users also need to install portaudio first. Additionally, install the websocket-client package:

    # (Mac)
    brew install portaudio

    # (Debian/Ubuntu)
    apt install portaudio19-dev

    pip install pyaudio
    pip install websocket-client

    In your code, first setup the microphone stream and then establish a WebSocket connection with the streaming endpoint by using a WebSocket client and connecting to wss://api.assemblyai.com/v2/realtime/ws.

    Authenticate your request by including your API key in the authorization header of your WebSocket connection, and provide the sample rate of your audio data as a query parameter to the streaming endpoint.

  2. 2

    Update the WebSocket's message event to load the incoming data as JSON and extract the text

  3. 3

    Update the WebSocket's message event to print the transcript, conditionally prepended with a string that signifies if the transcript is partial or final.

  4. 4

    Update the WebSocket's open event to stream data from the microphone.

  5. 5

    Optional: Add up to 2,500 characters of custom vocabulary to your streaming session by including the word_boost parameter as an optional query parameter in the URL.

    See also Adding Custom Vocabulary

  6. 6

    Update the WebSocket's error event to handle WebSocket errors and application-level errors, including bad sample rate, authentication failure, insufficient funds, and more. See also Closing and Status Codes for a list of errors.

    Additionally, update the WebSocket's close event.

Audio Requirements​

The raw audio data must comply with a strict encoding format. This is because we don't do any transcoding to your data, we send it directly to the model for transcription to reduce latency. The encoding of your audio must be in:

tip

Audio segments with a duration between 100 ms and 450 ms produce the best results in transcription accuracy.

Specifying the encoding​

By default, transcriptions expect PCM16 encoding. If you want to use mu-law encoding, you must set the encoding parameter to pcm_mulaw:

wss://api.assemblyai.com/v2/realtime/ws?sample_rate=16000?encoding=pcm_mulaw
EncodingDescription
pcm_s16le (Default)PCM signed 16-bit little-endian.
pcm_mulawPCM Mu-law.

Request Types​

These are the types of requests that can be sent to the WebSocket API.

Opening a Session​

When opening a Session you can pass the following query attributes to the WebSocket URL:

sample_rate​

The sample rate of the streamed audio.

Example: wss://api.assemblyai.com/v2/realtime/ws?sample_rate=16000

word_boost​

See also Adding Custom Vocabulary

encoding​

See also Specifying the encoding

token​

See also Creating Temporary Authentication Tokens

Sending Audio​

When sending audio over the WebSocket connection, you can use the websocket's binary mode to send raw audio data. This can be the raw data recorded directly from a microphone or read from an audio file.

# read from the microphone
data = stream.read(FRAMES_PER_BUFFER)

# binary data can be sent directly
ws.send(data)

# Note: Some WebSocket clients require that you specify the type:
# ws.send(data, opcode=websocket.ABNF.OPCODE_BINARY)
Heads up

Sending audio_data via JSON is also supported but will be deprecated in the future. Use the binary mode instead.

FieldExampleDescription
audio_data"UklGRtjIAABXQVZFZ"Raw audio data, base64 encoded.

Terminating a Session​

When you've completed your session, clients should send a JSON message with the following field.

FieldExampleDescription
terminate_sessiontrueA boolean value to communicate that you wish to end your streaming session forever.

After requesting session termination, the server will send the remaining transcript messages, followed by a SessionTerminated message.

Response Types​

These are the types of responses that can be received from the WebSocket API.

Session Start​

Once your request is authorized and connection established, your client receives a SessionBegins message with the following JSON data:

FieldExampleDescription
message_type"SessionBegins"Describes the type of the message.
session_id"d3e8c537-2f11-494b-b497-e59a434588bd"Unique identifier for the established session.
expires_at"2023-05-24T08:09:10.161850"Timestamp when this session will expire.

Transcripts​

Our Streaming Speech-to-Text pipeline uses a two-phase transcription strategy, broken into partial and final results.

Partial Transcripts​

As you send audio data to the API, the API immediately starts responding with Partial Results. The following keys are returned from the WebSocket API.

FieldExampleDescription
message_type"PartialTranscript"Describes the type of message.
audio_start0Start time of audio sample relative to session start, in milliseconds.
audio_end1500End time of audio sample relative to session start, in milliseconds.
confidence0.987190506414702The confidence score of the entire transcription, between 0 and 1.
text"there is a house in new orleans"The partial transcript for your audio.
words[{"start": 0, "end": 440, "confidence": 1.0, "text": "there"}, ...]An array of objects, with the information for each word in the transcription text. Includes the start/end time (in milliseconds) of the word, the confidence score of the word, and the text (i.e. the word itself).
created"2023-05-24T08:09:10.161850"The timestamp for the partial transcript.

Final Transcripts​

After you've received your partial results, our model continues to analyze incoming audio and, when it detects the end of an "utterance" (usually a pause in speech), it'll finalize the results sent to you so far with higher accuracy, as well as add punctuation and casing to the transcription text.

The following keys are returned from the WebSocket API when Final Results are sent:

FieldExampleDescription
message_type"FinalTranscript"Describes the type of message.
audio_start0Start time of audio sample relative to session start, in milliseconds.
audio_end1500End time of audio sample relative to session start, in milliseconds.
confidence0.997190506414702The confidence score of the entire transcription, between 0 and 1.
text"There is a house in New Orleans"The final transcript for your audio.
words[{"start": 0, "end": 440, "confidence": 1.0, "text": "There"}, ...]An array of objects, with the information for each word in the transcription text. Includes the start/end time (in milliseconds) of the word, the confidence score of the word, and the text (i.e. the word itself).
created"2023-05-24T08:09:10.161850"The timestamp for the final transcript.
punctuatedtrueWhether the text has been punctuated and cased.
text_formattedtrueWhether the text has been formatted (e.g. Dollar -> $)

Session Terminated​

After requesting session termination, the server will send the remaining transcript messages, followed by a SessionTerminated message. Your client receives a SessionTerminated message with the following JSON data:

FieldExampleDescription
message_type"SessionTerminated"Describes the type of the message.

Closing and Status Codes​

The WebSocket specification provides standard errors.

Our API provides application-level WebSocket errors for well-known scenarios:

Error ConditionStatus CodeMessage
bad sample rate4000"Sample rate must be a positive integer"
auth failed4001"Not Authorized"
insufficient funds4002"Insufficient Funds"
free tier user4003"This feature is paid-only and requires you to add a credit card. Please visit https://app.assemblyai.com/ to add a credit card to your account"
attempt to connect to nonexistent session id4004"Session not found"
session expired4008"Session Expired"
attempt to connect to closed session4010"Session previously closed"
rate limited4029"Client sent audio too fast"
unique session violation4030"Session is handled by another WebSocket"
session times out4031"Session idle for too long"
audio too short4032"Audio duration is too short"
audio too long4033"Audio duration is too long"
audio too small to transcode4034"Audio too small to transcode"
bad schema4101"Endpoint received a message with an invalid schema"
too many streams4102"This account has exceeded the number of allowed streams"
reconnected4103"This session has been reconnected. This WebSocket is no longer valid"
word boost parameter parsing failed4104"Could not parse word boost parameter"

Quotas and Limits​

The following limits are imposed to ensure performance and service quality.

  • Idle Sessions - Sessions that don't receive audio within 1 minute will be terminated.
  • Session Limit - 100 sessions at a time for paid users. Please contact us if you need to increase this limit. Free-tier users must upgrade their account to use real-time streaming.
  • Session Uniqueness - Only one WebSocket per session.
  • Audio Sampling Rate Limit - Customers must send data in near real-time. If a client sends data faster than 1 second of audio per second for longer than 1 minute, we'll terminate the session.

Adding Custom Vocabulary​

Developers can also add up to 2500 characters of custom vocabulary to their real-time session by adding the optional query parameter word_boost in the URL. The parameter should map to a JSON encoded list of strings as shown in this Python example:

import json
from urllib.parse import urlencode

sample_rate = 16000
word_boost = ["foo", "bar"]
params = {"sample_rate": sample_rate, "word_boost": json.dumps(word_boost)}

url = f"wss://api.assemblyai.com/v2/realtime/ws?{urlencode(params)}"

Creating Temporary Authentication Tokens​

If you need to authenticate on the client, you can avoid exposing your API key by using temporary authentication tokens. Temporary tokens have a one-time use restriction. To generate a temporary token, send a POST request to https://api.assemblyai.com/v2/realtime/token. Use the expires_in parameter to specify how long the token should be valid for, in seconds.

note

The expires_in parameter must have a value between 60 and 360000 seconds.

curl --request POST \
  --url https://api.assemblyai.com/v2/realtime/token \
  --header 'authorization: YOUR_API_KEY' \ 
  --header 'Content-Type: application/json' \
  --data '{"expires_in": 60}'

In response you'll receive the following JSON output:

{
  "token": "b2e3c6c71d450589b2f4f0bb1ac4efd2d5e55b1f926e552e02fc0cc070eaedbd"
}

A developer can now use this temporary token in the browser to authenticate a new WebSocket session with the following endpoint wss://api.assemblyai.com/v2/realtime/ws?sample_rate=16000&token={New Temp Token}. For example:

let socket
const token = 'b2e3c6c71d450589b2f4f0bb1ac4efd2d5e55b1f926e552e02fc0cc070eaedbd'

socket = new WebSocket(
  `wss://api.assemblyai.com/v2/realtime/ws?sample_rate=16000&token=${token}`
)

Conclusion​

Streaming Speech-to-Text is a powerful feature with even more powerful possibilities for integration. On the AssemblyAI blog, you can learn about using Streaming Speech-to-Text to:

You can also find an example of using Express.js for Streaming Speech-to-Text on GitHub.

© AssemblyAI