Deepgram Pricing & Developer Hub

Deepgram Speech Intelligence: Nova-2 Transcription & Aura TTS Economics

Deepgram is the industry leader in enterprise speech AI, delivering end-to-end voice processing across **Nova-2 Speech-to-Text (STT)** and **Aura Text-to-Speech (TTS)**. Built on proprietary deep learning architectures rather than standard open-source Whisper checkpoints, Nova-2 achieves an ultra-low Word Error Rate (WER) with streaming latencies under **250ms**. At **$0.0043 per audio minute ($0.258/hour)** for streaming, Deepgram is up to **75% cheaper than Google Cloud STT ($0.016/min)** and AWS Transcribe ($0.024/min).

Deepgram Speech AI Production Rate Card

Deepgram Streaming STT Client with Interim Results (Node.js)

import { createClient, LiveTranscriptionEvents } from '@deepgram/sdk';

const deepgram = createClient(process.env.DEEPGRAM_API_KEY);

export function initializeLiveTranscription(audioStream: NodeJS.ReadableStream) {
  const connection = deepgram.listen.live({
    model: 'nova-2',
    smart_format: true,
    interim_results: true,
    endpointing: 300 // Trigger final transcript after 300ms of user silence
  });

  connection.on(LiveTranscriptionEvents.Transcript, (data) => {
    const text = data.channel.alternatives[0].transcript;
    if (text) console.log(`[${data.is_final ? 'FINAL' : 'INTERIM'}]:`, text);
  });

  audioStream.on('data', (chunk) => connection.send(chunk));
}

Engineering Q&A: Deepgram Voice Architecture

Acoustic Model Training & Custom Keyword Vocabulary Boosting

While generic speech APIs struggle with specialized corporate jargon, brand names, and medical terms, Deepgram Nova-2 provides native **Keyword Boosting**. By passing an array of proprietary terms (e.g. `keywords: ["CheckApiCost:3.5", "Kubernetes:2.0"]`), Nova-2 biases its acoustic phoneme decoding lattice toward those specific phrases with zero extra transcription latency or custom training fees, increasing technical transcription precision by up to 40%.

Deepgram Production SRE Checklist

• Audio Encoding Optimization: For high-volume telephony pipelines, stream linear PCM16 mono audio at 16,000 Hz. Deepgram's neural acoustic front-end is natively trained on 16kHz audio, delivering optimal accuracy with minimal decoding latency.
• Endpointing Timeout Calibration: Set `endpointing: 250` (250 milliseconds). In conversational phone bots, waiting 1,000ms for a customer to stop speaking creates awkward pauses. Tuning endpointing to 250ms triggers immediate bot turn-taking responses.

Deepgram Multi-Channel Audio Processing & Telephony Diarization

In professional contact center environments, calls are typically recorded as dual-channel stereo audio (agent on channel 0, caller on channel 1). Deepgram Nova-2 supports native multi-channel transcription by setting `multichannel=true` in the WebSocket configuration. Transcribing both tracks simultaneously allows Deepgram to assign words to exact speaker channels with 100% acoustic certainty, completely bypassing the computational overhead and accuracy degradation of probabilistic diarization algorithms.

Acoustic Latency Optimization: Enable Deepgram Smart Format to automatically parse currency figures, calendar dates, and telephone numbers into natural spoken text formatting, eliminating the need for post-processing regex scripts.