AWS

Use Amazon Transcribe (STT), Amazon Bedrock (LLM), and Amazon Polly (TTS) with Voxray. All three services are covered by the standard AWS SDK v2 credential chain — no separate API key is required.

STT

Amazon Transcribe — real-time streaming transcription

LLM

Amazon Bedrock — Claude, Titan, Llama, and more

TTS

Amazon Polly — neural voices in 30+ languages

Overview

AWS is unique among Voxray providers: it uses the AWS SDK v2 credential chain rather than a single bearer token. All three pipeline stages — speech recognition, language model inference, and speech synthesis — resolve credentials the same way and share a single region setting.

Set "stt_provider", "llm_provider", and "tts_provider" to "aws" in your config to activate all three, or mix AWS with other providers as needed.

Authentication

Environment Variables

The simplest method for local development and CI pipelines. Set the following before starting Voxray:

export AWS_ACCESS_KEY_ID=AKIAIOSFODNN7EXAMPLE
export AWS_SECRET_ACCESS_KEY=wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY
# Optional — required for temporary credentials from STS
export AWS_SESSION_TOKEN=AQoDYXdzEJr...

The SDK picks these up automatically; no config change is needed.

~/.aws/credentials

The shared credentials file is the standard method for developer machines. Run aws configure or edit the file directly:

[default]
aws_access_key_id     = AKIAIOSFODNN7EXAMPLE
aws_secret_access_key = wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY

To use a named profile, set AWS_PROFILE=my-profile before starting Voxray. The SDK reads ~/.aws/credentials and ~/.aws/config automatically.

IAM Role (Recommended for Production)

When Voxray runs on EC2, ECS, Lambda, or any service that supports instance metadata, the SDK fetches credentials from the Instance Metadata Service (IMDS) with no configuration required.

Attach an IAM role to your compute resource with the permissions listed in the Required IAM Permissions section below. The SDK refreshes temporary credentials before they expire.

This is the recommended approach for production because no long-lived secrets need to be stored or rotated manually.

Region Configuration

All AWS services require a region. Voxray resolves the region in this order:

1. aws_region key inside api_keys in your config file
2. AWS_REGION environment variable
3. Default: us-east-1

Via config file:

{
  "api_keys": {
    "aws_region": "us-west-2"
  }
}

Via environment variable:

export AWS_REGION=eu-west-1

Choose a region where all three services — Transcribe, Bedrock, and Polly — are available and where you have enabled the Bedrock models you intend to use.

Quick Start Config

The minimal configuration to run a fully AWS-native voice pipeline:

{
  "stt_provider": "aws",
  "llm_provider": "aws",
  "tts_provider": "aws",
  "model": "anthropic.claude-3-haiku-20240307-v1:0",
  "api_keys": {
    "aws_region": "us-east-1"
  }
}

No aws key is needed inside api_keys — credentials come from the SDK credential chain described above. The aws key in api_keys is accepted for legacy compatibility but is not used for authentication.

Amazon Bedrock (LLM)

Amazon Bedrock provides access to foundation models from Anthropic, Amazon, Meta, and others via a unified API. You must enable model access in the Bedrock console for each model ID you use; models are not accessible by default.

Supported Models

Model ID	Family	Notes
anthropic.claude-3-haiku-20240307-v1:0	Anthropic Claude 3	Fastest Claude 3; lowest latency
anthropic.claude-3-sonnet-20240229-v1:0	Anthropic Claude 3	Balanced quality and speed
anthropic.claude-3-5-sonnet-20240620-v1:0	Anthropic Claude 3.5	Highest quality Claude on Bedrock
amazon.titan-text-express-v1	Amazon Titan	Amazon’s own foundation model
meta.llama3-8b-instruct-v1:0	Meta Llama 3	Open-weight 8B instruction model
meta.llama3-70b-instruct-v1:0	Meta Llama 3	Open-weight 70B instruction model

Set the model in your config:

{
  "llm_provider": "aws",
  "model": "anthropic.claude-3-haiku-20240307-v1:0"
}

Model availability varies by region. Check the Bedrock console in your target region to confirm a model is available and request access before deploying.

Amazon Transcribe (STT)

Amazon Transcribe streams audio in real time and returns partial and final transcription results. Voxray uses the StartStreamTranscription API over a bidirectional WebSocket — no intermediate file uploads are required.

Key characteristics:

• Real-time streaming transcription with incremental results
• Supports 30+ languages including en-US, en-GB, es-US, fr-FR, de-DE, ja-JP, hi-IN, pt-BR, and more
• Automatic punctuation and number normalization
• No per-request cold start once the stream is open

To activate:

{
  "stt_provider": "aws"
}

The language is inferred from the audio stream. To pin a specific language, use the stt_language config key (e.g. "en-US").

Amazon Polly (TTS)

Amazon Polly synthesizes text to speech using neural voices. Voxray streams the audio output directly into the pipeline without buffering the full response.

Key characteristics:

• Neural TTS engine with natural-sounding voices
• 30+ languages and 60+ voices
• Low-latency synthesis suitable for real-time voice agents

Voice Selection

Set the voice ID via tts_voice in your config. If no voice is set, Voxray defaults to Joanna.

{
  "tts_provider": "aws",
  "tts_voice": "Matthew"
}

Common Voice IDs

Voice ID	Language	Gender
Joanna	en-US	Female
Matthew	en-US	Male
Amy	en-GB	Female
Brian	en-GB	Male
Celine	fr-FR	Female
Vicki	de-DE	Female
Mizuki	ja-JP	Female
Lucia	es-ES	Female
Camila	pt-BR	Female

For a full list, see the Amazon Polly documentation.

Required IAM Permissions

Attach these permissions to the IAM role or user that Voxray runs as. The principle of least privilege is applied — only the specific actions each service requires are listed.

Permission	Service	Purpose
transcribe:StartStreamTranscription	Amazon Transcribe	Open a real-time transcription stream
bedrock:InvokeModelWithResponseStream	Amazon Bedrock	Streaming LLM inference
polly:SynthesizeSpeech	Amazon Polly	Neural TTS synthesis

Minimal IAM policy document:

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Sid": "VoxrayTranscribe",
      "Effect": "Allow",
      "Action": ["transcribe:StartStreamTranscription"],
      "Resource": "*"
    },
    {
      "Sid": "VoxrayBedrock",
      "Effect": "Allow",
      "Action": ["bedrock:InvokeModelWithResponseStream"],
      "Resource": "arn:aws:bedrock:*::foundation-model/*"
    },
    {
      "Sid": "VoxrayPolly",
      "Effect": "Allow",
      "Action": ["polly:SynthesizeSpeech"],
      "Resource": "*"
    }
  ]
}

If you also use Voxray’s S3 conversation recording, add s3:PutObject on your recording bucket to the same policy.

Full Example Config

{
  "host": "0.0.0.0",
  "port": 8080,
  "transport": "both",

  "stt_provider": "aws",
  "llm_provider": "aws",
  "tts_provider": "aws",

  "model": "anthropic.claude-3-haiku-20240307-v1:0",
  "tts_voice": "Joanna",

  "api_keys": {
    "aws_region": "us-east-1"
  },

  "webrtc_ice_servers": [
    "stun:stun.l.google.com:19302"
  ]
}

Run with credentials resolved via IAM role (EC2/ECS) or environment variables (local development):

./voxray -config config.json