Deepgram Nova-3 is an advanced speech-to-text model built on Deepgram's end-to-end deep learning architecture. It delivers industry-leading word error rates for English, real-time streaming transcription via WebSocket, and a rich set of audio intelligence features — making it suitable for contact centers, meeting transcription, voice analytics, and developer applications.
Technical Specifications
Performance Benchmarks
- Industry-leading word error rate (WER) for English.
- Sub-300ms latency in streaming mode.
- Handles noisy environments, accents, and spontaneous speech.
- Supports audio files up to 2GB or 5 hours in batch mode.
- Per-word confidence scores available in output.
Architecture Breakdown
Nova-3 uses Deepgram's proprietary end-to-end deep learning pipeline that processes raw audio directly, eliminating multi-step feature extraction. This reduces error accumulation and enables faster, more accurate inference across diverse recording conditions.
Pricing
- $0.01001 / min
Core Features & Capabilities
- Streaming Transcription: Real-time interim and final results via WebSocket.
- Speaker Diarization: Automatically labels each speaker starting at index 0.
- Smart Formatting: Locale-aware formatting for numbers, dates, and punctuation.
- Intent & Topic Detection: Detects custom or model-identified intents and topics.
- Entity Detection: Extracts key entities from audio content.
- Custom Vocabulary (keyterm): Boosts recognition accuracy for domain-specific terms.
- Filler Word Detection: Identifies and optionally removes filler words.
- Utterance Segmentation: Splits transcript into labeled speech segments.
- Multichannel Support: Transcribes each audio channel independently.
Comparison with Other Models
VS Deepgram Nova-3 General: Nova-3 is optimized for English-only workflows with maximum accuracy, while Nova-3 General adds multilingual support at the same price point.
VS Deepgram Nova-3 Medical: Nova-3 Medical is fine-tuned for clinical terminology and healthcare audio; Nova-3 is the general-purpose English model at a higher per-minute rate.
Measured on the AI/ML API gateway across live traffic. Figures are medians unless stated. Last updated 2026-10-12.
Nova-3 is documented as suited to real-time contact center transcription and automation. It is described as fitting contact center workflows, including real-time transcription and automation with customer support systems.
Nova-3 is documented as suited to meeting transcription and event captioning. The model is listed as recommended for meetings and event captioning, including multi-speaker audio.
Nova-3 is documented as suited to noisy or far-field audio. Deepgram describes it as a strong fit for noisy, multi-speaker, and far-field environments such as drive-thrus and call centers.
Nova-3 is documented as suited to multilingual live applications. Deepgram says it supports real-time multilingual transcription and is suited to global customer support, voice agents, and live captioning.
import requests, time headers = {"Authorization": "Bearer " + AIMLAPI_KEY} job = requests.post( "https://api.aimlapi.com/v1/stt/create", headers=headers, json={ "model": "deepgram/nova-3", "url": "https://example.com/audio.mp3" }, ).json() gid = job["generation_id"] while True: res = requests.get(f"https://api.aimlapi.com/v1/stt/{gid}", headers=headers).json() if res.get("status") in ("completed", "error", "failed"): break time.sleep(3) print(res)
const headers = { Authorization: `Bearer ${process.env.AIMLAPI_KEY}`, "Content-Type": "application/json", }; const job = await (await fetch("https://api.aimlapi.com/v1/stt/create", { method: "POST", headers, body: JSON.stringify({ "model": "deepgram/nova-3", "url": "https://example.com/audio.mp3" }), })).json(); let res; do { await new Promise((r) => setTimeout(r, 3000)); res = await (await fetch(`https://api.aimlapi.com/v1/stt/${job.generation_id}`, { headers })).json(); } while (!["completed", "error", "failed"].includes(res.status)); console.log(res);
# submit the job — the response contains "generation_id" curl -X POST https://api.aimlapi.com/v1/stt/create \ -H "Authorization: Bearer $AIMLAPI_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"deepgram/nova-3","url":"https://example.com/audio.mp3"}' # then poll for the result until it is ready curl "https://api.aimlapi.com/v1/stt/{generation_id}" -H "Authorization: Bearer $AIMLAPI_KEY"
OpenAI-compatible — swap the base URL and it works with your existing SDK.
| Type | Price |
|---|---|
| Output | |
| Model | Input | Output | Context | Best for |
|---|---|---|---|---|
Nova-3 This page | Transcription | |||
| Speech synthesis | ||||
| Speech synthesis | ||||
| Speech synthesis | ||||
| Speech synthesis |
Nova-3 takes audio as input and returns text.
Nova-3 became available on June 4, 2026.
Nova-3 is priced at output $0.0001668 / sec.
Nova-3 was built by Deepgram.
Use deepgram/nova-3 as the model id on AI/ML API.
Yes. Nova-3 is served through AI/ML API, so the same key and endpoint format used for other models applies.
Yes, Nova-3 supports streaming as one of its capabilities, alongside reasoning.
