Bandwidth Speech-to-Text Model

AI model · Speech to text · Audio / speech → Text by Bandwidth

Ultra-low latency speech-to-text built for the voice network.

A proprietary speech-to-text model built for ultra-low latency telephony. It runs on edge GPUs within Bandwidth's network to provide real-time transcription and PII redaction without third-party cloud hops.

Model specs

The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.

Languages
English

Category: Natural Language Processing

Modalities
Audio (Telephony) → Text
Real-time streaming
Real-time streaming-native processing
Latency
Ultra-low latency with no extra cloud hops
Access
API integration via Bandwidth Universal Platform
Input modality
Audio (Telephony)
Output modality
Text
  • Ultra-low latency streaming for telephony
  • Runs on proprietary edge GPUs in the network
  • Streaming-native with no word revisions
  • Real-time PII and credit card redaction
  • Eliminates third-party cloud hops

Overview

Bandwidth Speech-to-Text Model is a proprietary AI model built for ultra-low latency streaming in real-world telephony. It is designed for businesses and developers who need to convert live voice conversations into text with minimal delay for applications like virtual agents and call analytics. The model runs on Bandwidth's own edge GPUs, which eliminates the need for extra hops to third-party clouds and ensures a direct, high-performance voice path. It features streaming-native processing that avoids word revisions and includes built-in capabilities for real-time PII redaction. This model is part of Bandwidth's Super Intelligence initiative to bring AI processing closer to the network edge. It is delivered as part of the Bandwidth Call Intelligence suite on the Universal Platform.

Performance and Architecture

  • Ultra-low latency streaming: The model is built from the ground up specifically for ultra-low latency streaming in real-world telephony environments.
  • Edge GPU inference stack: The model is built and run on Bandwidth's own edge GPUs, eliminating extra hops to third-party clouds for faster processing.
  • Streaming-native processing: The model is streaming-native, providing steady results without word revisions during the call.
  • Real-time PII redaction: The system can identify and redact sensitive information like credit card numbers in real-time as they are being spoken.

Developer Features

  • Telephony optimized: The model is designed to handle the specific audio characteristics and noise profiles of real-world telephony.
  • API integration: Developers can create an API key to integrate the model directly into their voice applications and workflows.
  • Live audio support: The model supports live microphone audio for testing and real-time transcription demonstrations.

Use cases

  • Voice AI builders use the model to create virtual agents that respond with minimal latency during live calls.
  • Contact centers implement real-time transcription to power live call analytics and agent assistance tools.
  • Regulated industries use the built-in redaction to protect sensitive customer data like credit card numbers during live sessions.

Specifications

Modalities
Audio (Telephony) → Text
Real-time streaming
Real-time streaming-native processing
Latency
Ultra-low latency with no extra cloud hops
Access
API integration via Bandwidth Universal Platform
Input modality
Audio (Telephony)
Output modality
Text
Inference location
Edge GPUs on Bandwidth network
Latency type
Ultra-low latency streaming
Processing type
Streaming-native (no revisions)

How it works

  1. API Integration: Developers create an API key through the Bandwidth Universal Platform to access the model's capabilities.
  2. Edge Processing: Voice calls are processed directly on Bandwidth's edge GPUs co-located within the voice network path.
  3. Real-time Transcription: The model converts live audio to text in real-time, optionally applying PII redaction as words are spoken.

FAQ

How does the model achieve ultra-low latency?

The model runs on Bandwidth's own edge GPUs co-located where the phone call happens, eliminating the extra hops to third-party clouds that typically add latency.

What does streaming-native mean for transcription?

Streaming-native means the model provides steady results without word revisions during the call, which is critical for real-time applications like virtual agents.

Can the model redact sensitive information?

Yes, the model includes built-in capabilities to redact sensitive information, such as credit card numbers, in real-time as they are spoken.

Is this model part of a larger product?

The model is part of Bandwidth's Call Intelligence solution, which unifies call recording, transcription, and redaction in one step.

Added by AntSpace from public sources (labs.bandwidth.com).

More AI models by Bandwidth