How To Deploy Smartbot V2

Version Summary

Release Notes

SmartBot V2 introduces a multimodal sentiment analysis model capable of processing raw audio inputs in real time. This version supports base64-encoded .wav inputs and performs speech-to-text transcription, spectrogram generation, and text-audio fusion for sentiment analysis. Compared to V1, this version has improved accuracy on diverse speakers and output classes.

Model Input Details

SmartBot V2 accepts JSON input structured as a dictionary of base64-encoded .wav audio clips. Each entry must include a unique uid key. No text input is required — the system performs speech-to-text internally.

Input Constraints

• Audio format: mono .wav, 16-bit PCM, 16kHz
• Max file size per clip: 3 MB
• Max clips per request: 10

Input Format

{
  "invocations": {
    "uid123": "<base64-encoded-audio>",
    "uid456": "<base64-encoded-audio>"
  }
}

Real-Time Input Sample

{
  "invocations": {
    "12345": "UklGRtQAAABXQVZFZm10IBAAAAABAAEA...",
    "jh13d3": "UklGRkQAAABXQVZFZm10IBAAAAABAAEA..."
  }
}

Model Output Details

Returns a list of dictionaries, each containing:

• uid — original identifier
• predicted_class — numeric class index
• predicted_label — predicted class label (happy, neutral, sad, anger)
• probabilities — full distribution across all output classes

Output Sample

{
  "predictions": [
    {
      "uid": "12345",
      "predicted_class": 2,
      "predicted_label": "neutral",
      "probabilities": {
        "anger": 0.0094,
        "happy": 0.0029,
        "neutral": 0.5295,
        "sad": 0.4580
      }
    },
    {
      "uid": "jh13d3",
      "predicted_class": 1,
      "predicted_label": "happy",
      "probabilities": {
        "anger": 0.0030,
        "happy": 0.9451,
        "neutral": 0.0212,
        "sad": 0.0307
      }
    }
  ]
}

Usage Instructions

• Prepare .wav files using standard tools (Audacity, FFmpeg, etc.).
• Convert to base64 and pass as JSON under the "invocations" field.
• The endpoint handles STT + spectrogram generation internally.
• Output includes sentiment predictions and class probabilities.

Authentication and Endpoint Setup

Subscribe to the Smartbot V2 model package via AWS Marketplace. On subscription, the model package ARN is automatically shared to your AWS account.

Ensure the IAM role used to deploy the endpoint has the following permissions:

• sagemaker:CreateModel
• sagemaker:CreateEndpointConfig
• sagemaker:CreateEndpoint
• sagemaker:InvokeEndpoint

Deploy the endpoint using the AWS Console, SDK, or CLI. Example using boto3:

import boto3

sm = boto3.client("sagemaker", region_name="eu-west-2")

sm.create_model(
  ModelName="smartbot-v2",
  PrimaryContainer={
    "ModelPackageName": "arn:aws:sagemaker:eu-west-2:923927835297:model-package/smartbot-v2-july-version-2025"
  },
  ExecutionRoleArn="arn:aws:iam::<your-account-id>:role/<your-sagemaker-role>"
)

Once the endpoint is in-service, invoke it using sagemaker-runtime:

runtime = boto3.client("sagemaker-runtime", region_name="eu-west-2")

response = runtime.invoke_endpoint(
  EndpointName="smartbot-v2",
  ContentType="application/json",
  Body=json.dumps(payload)
)

The full working example is available in the sample notebook linked in the Resources section below.

Latency

Typical real-time inference latency on ml.m5.xlarge:

For contact centre deployments requiring sub-second sentiment signals, invoke the endpoint asynchronously in parallel with the live call rather than inline in the call path.

To avoid cold start latency in production, use SageMaker endpoint auto-scaling or a keep-warm strategy.

Recommended Instance Types

Resources