GuidesVoice providers
Deepgram in Australia: pricing, Nova-3, Flux and the Sydney endpoint
Call 1800 570 693 to talk to the demo agent. Plans start at A$299 a month, GST included.
- Australian endpoint
- api.au.deepgram.com, in AWS Sydney
- Streaming speech to text
- USD 0.0077 a minute list
- Voice agent API
- USD 0.075 a minute
- Free credit
- USD 200, no card needed
Read on deepgram.com and developers.deepgram.com on 6 October 2026. Pay-as-you-go prices in US dollars, GST not included.
Deepgram models for voice agents
Deepgram makes two families of speech to text. Nova-3 transcribes in more than 60 languages, including Australian and New Zealand English. Flux is built for voice agents and decides when the caller has finished speaking.
Aura-2 is Deepgram's text to speech in seven languages, and Flux TTS, launched on 12 August 2026, is a voice model tuned for conversation. The Voice Agent API joins speech to text, a language model and text to speech behind one connection.
| Model | Use | Languages | Deepgram's latency claim |
|---|---|---|---|
| Nova-3 | Speech to text, streaming and files | 60+, en-AU included | Not published as one figure |
| Flux | Speech to text with turn detection | English, or 10 languages on flux-general-multi | About 260 ms to detect the end of a turn |
| Nova-3 Medical | Clinical speech to text | 8 English variants, en-AU included | Not published |
| Aura-2 | Text to speech | 7 | About 90 ms to first audio |
| Flux TTS | Text to speech for conversation | Not published | As little as 80 ms to first audio |
| Voice Agent API | Speech to text, model and voice in one | Follows the models chosen | Under 300 ms |
Deepgram pricing
Deepgram bills by the minute of audio for speech to text and by the character for text to speech. The pricing page shows a promotional rate beside the list rate for streaming speech to text, with no end date. Budget on the list rate.
Growth plans start at USD 4,000 a year of prepaid credit and cut the rates by up to 20 per cent. Enterprise is quoted.
| Product | Pay as you go | Note |
|---|---|---|
| Nova-3 streaming, one language | USD 0.0077 a minute | Promotional USD 0.0048 |
| Nova-3 streaming, multilingual | USD 0.0092 a minute | Promotional USD 0.0058 |
| Flux streaming, English | USD 0.0077 a minute | Promotional USD 0.0065 |
| Nova-3 pre-recorded | USD 0.0043 a minute | Multilingual USD 0.0052 |
| Aura-2 text to speech | USD 0.030 per 1,000 characters | Aura-1 USD 0.015 |
| Flux TTS | USD 0.045 per 1,000 characters | Launch credit to 31 December 2026 |
| Voice Agent API, standard | USD 0.075 a minute | USD 0.065 with your own voice |
| Voice Agent API, advanced | USD 0.163 a minute | USD 0.122 with your own voice |
Cost per call minute
A phone agent transcribes every minute the caller is on the line and speaks for part of it. At 900 characters a spoken minute, Aura-2 costs about USD 0.027 per minute of agent speech. Nova-3 streaming adds USD 0.0077 per minute of call.
Redaction, speaker labels and keyterm prompting each cost extra on streaming, from USD 0.0013 to USD 0.0020 a minute.
The Sydney endpoint
Deepgram made api.au.deepgram.com generally available on 17 June 2026, hosted in AWS ap-southeast-2. It serves speech to text, text to speech, the Voice Agent API and Text Intelligence, with the same models as the European endpoint and no waitlist or Enterprise requirement.
Deepgram never reroutes a request sent to the Australian endpoint. If Sydney is down, the request fails, so a production agent needs a fallback plan for that case.
The Projects management API is global, and Deepgram keeps billing records in the United States. Our platform can route an agent's speech recognition to this endpoint, as our sub-processors page lists.
The model improvement opt-out
By default, Deepgram keeps audio and transcripts to improve its models, and its docs say that data may be processed outside your region. Setting mip_opt_out=true on each request stops this, and Deepgram then stores nothing after it responds.
Deepgram's own docs say the regional endpoint and the opt-out together keep data in Australia. Its compliance page ties its Australian Privacy Principle 8 statement to the opt-out being on.
- Send requests to api.au.deepgram.com, never api.deepgram.com.
- Add mip_opt_out=true to every request.
- Confirm in writing whether the Voice Agent API's language model step leaves Australia for the model you pick.
Flux turn detection
Flux sends events as the caller talks: StartOfTurn, Update, EagerEndOfTurn, TurnResumed and EndOfTurn. An agent can start preparing its reply at EagerEndOfTurn and speak it at EndOfTurn, which Deepgram says arrives 150 to 250 ms later.
Deepgram's Flux quickstart says the eager event raises language model calls by 50 to 70 per cent, because some turns resume. Flux runs only over the WebSocket endpoint at /v2/listen, and accepts mulaw at 8 kHz, the format most phone carriers send.
Running Deepgram in your own cloud
Deepgram's self-hosted product runs on Docker, Podman or Kubernetes in AWS, Azure, Google Cloud, Oracle Cloud or your own data centre. It needs an Enterprise licence and a dedicated NVIDIA GPU with at least 16 GB of memory; Deepgram recommends the L4.
A self-hosted install must reach license.deepgram.com on port 443 at all times, directly or through Deepgram's licence proxy. We found no fully offline option on Deepgram's pages.
Deepgram also sells Nova-3, Flux, Aura-2 and Flux TTS as Amazon SageMaker model packages that run in your own VPC, with live streaming over a bidirectional SageMaker stream. Deepgram recommends ml.g6.2xlarge and offers a 14-day trial. Ask aws@deepgram.com whether the packages run in ap-southeast-2.
Deepgram concurrency limits
Deepgram sets a limit on concurrent connections for each project. On the Australian endpoint, Flux TTS allows five concurrent streams below Enterprise, which caps a phone agent at five callers hearing it at once.
| Product | Pay as you go | Growth | Enterprise |
|---|---|---|---|
| Nova-3 and Flux streaming | 150 | 150 | 300 |
| Pre-recorded | 50 | 50 | 200+ |
| Aura-2 | 45 | 45 | 100 |
| Flux TTS | 5 | 5 | 50 |
| Voice Agent API | 45 | 45 | 100+ |
Deepgram compliance
Deepgram lists SOC 2 Type 1 and Type 2, GDPR, CCPA and PCI compliance on its data privacy page, and signs a HIPAA business associate agreement on request. We found no ISO 27001 certificate on a Deepgram page.
Questions
Deepgram runs api.au.deepgram.com in AWS Sydney for speech to text, text to speech and its Voice Agent API. Any plan can use it at the global price. Add mip_opt_out=true to each request, or Deepgram may keep audio and process it outside Australia to train its models.
Nova-3 streaming lists at USD 0.0077 a minute, shown at a promotional USD 0.0048. Aura-2 text to speech costs USD 0.030 per 1,000 characters and the Voice Agent API costs USD 0.075 a minute. Prices read on deepgram.com on 6 October 2026, in US dollars.
New accounts get USD 200 of credit with no card required. At the Nova-3 list rate, that covers about 26,000 minutes of streaming transcription.
Nova-3 transcribes speech in more than 60 languages. Flux transcribes English, or 10 languages on its multilingual model, and also tells your agent when the caller has finished speaking, with events such as EagerEndOfTurn and EndOfTurn.
Deepgram's self-hosted product runs on your own GPUs in AWS Sydney or a data centre on an Enterprise licence. The install has to reach license.deepgram.com to check its licence, and only usage metadata leaves.
Nova-3 lists Australian English (en-AU) as a supported language, and Nova-3 Medical includes it too. Flux's English model does not list en-AU separately.
Deepgram signs a HIPAA business associate agreement on request and holds SOC 2 Type 2. For Australian health data, pair the Sydney endpoint with the model improvement opt-out.
Sources
Related pages
Get started
Tell us what your callers ring about, and we build and test the agent for you.