Amazon Web Services Adds Native Deepgram Speech Recognition to Amazon Lex V2

Amazon Web Services has introduced native configuration support that allows Amazon Lex V2 conversational bots to use Deepgram for automated speech recognition. The update enables development teams to route live bot audio streams directly to Deepgram API endpoints using credentials stored in AWS Secrets Manager, with support for custom model identifiers across all AWS regions where Lex V2 operates.

Direct Audio Routing Between Amazon Lex V2 and Deepgram Endpoints

Amazon Lex V2 provides conversational artificial intelligence interfaces using voice and text. Automated speech recognition serves as the ingestion layer for voice-enabled bots, converting spoken input into text before natural language understanding models process intent.

Prior to native integration, sending live voice streams from Lex V2 bots to external speech recognition engines required intermediate serverless functions or custom middleware. These architectures required developers to capture raw audio payloads, coordinate streaming handoffs, and manage network retries manually.

With native configuration, Amazon Lex V2 manages the audio pipeline directly. Incoming voice traffic is streamed straight from the Lex V2 runtime to Deepgram endpoints for real-time transcription. Lex V2 automatically directs traffic based on data locality:

• European AWS Regions: Bots operating in regions with the eu- prefix (such as eu-west-1, eu-west-2, and eu-central-1) automatically route requests to api.eu.deepgram.com.
• Global AWS Regions: Bots in all other supported AWS regions route requests to api.deepgram.com.

Credential Management and Symmetric KMS Encryption

Authorization between Amazon Lex V2 and Deepgram relies on AWS Secrets Manager, eliminating the need to hardcode API keys in bot definitions or application logic.

To configure access, administrators must store their Deepgram API token in Secrets Manager using a specific key-value pair and cryptographic setup:

1. Create a Customer Managed Key: A symmetric AWS Key Management Service (AWS KMS) key must be created to encrypt the secret. The default AWS-managed KMS key is not supported for this integration.
2. Store the Secret: Inside Secrets Manager, administrators configure an “Other type of secret” containing a single key-value entry:

   {
     "apiToken": "YOUR_DEEPGRAM_API_KEY"
   }

3. Grant IAM Permissions: The Lex V2 bot execution role requires permissions to read the secret and decrypt it using the associated KMS key.

Key updates made within AWS Secrets Manager apply dynamically at runtime, allowing key rotation without redeploying the bot.

Custom Model Identifiers

When configuring speech recognition in Lex V2, administrators can select Deepgram under the Speech model preference settings for any specific bot locale.

The configuration interface includes an optional Model ID field:

• Default Processing: If the field is left blank, Lex V2 defaults to Deepgram’s standard baseline transcription model.
• Custom Models: Teams can input specific Deepgram model identifiers to route speech recognition through specialized or fine-tuned domain models, such as models optimized for finance, healthcare, or custom vocabulary lists.

Regional Availability and Configuration Interfaces

Native Deepgram integration is available across all AWS commercial regions supporting Amazon Lex V2. The integration can be provisioned through the AWS Management Console, the AWS Command Line Interface (CLI), and AWS software development kits (SDKs).