ElevenLabs vs Cartesia vs Deepgram, and the other voice AI providers in Australia

ElevenLabs, Cartesia and Deepgram side by side

Each of the three sells text to speech, speech to text and a hosted voice agent, and each bills in US dollars. They differ most on where the audio is processed and on what you can run in your own cloud.

Providers that process speech in Australia

These providers publish an Australian region for speech. Billing records, support access and model training settings can sit offshore when the audio does not, so check each one.

• Speech to text, text to speech and its voice agent in AWS Sydney at api.au.deepgram.com, on any plan. Needs mip_opt_out=true.Speech to textVoiceSydney
• A dedicated Australian deployment on Enterprise, and Sonic 3 on SageMaker in your own AWS account in Sydney.VoiceSpeech to textEnterprise
• Microsoft Azure AI SpeechReal-time speech to text at USD 1.00 an hour and neural voices at USD 15 per million characters in australiaeast, with 15 en-AU voices.Speech to textVoiceaustraliaeast
• Amazon Polly and TranscribePolly neural voices at USD 16 per million characters and Transcribe streaming at USD 0.01 a minute, both in ap-southeast-2.Speech to textVoiceap-southeast-2
• Google Cloud Text-to-SpeechChirp 3 HD voices at USD 30 per million characters, callable in australia-southeast1. Chirp 3 speech to text runs in the US and EU only.Voiceaustralia-southeast1
• Real-time speech to text on an Australian endpoint, au.rt.speechmatics.com, from USD 0.30 an hour on its Linden model.Speech to textAustralia

Other providers voice teams use

These providers publish no Australian region, so speech sent to them is processed overseas.

• The widest language range of the three, 90+ on v4. Residency in the EU, India and Singapore on Enterprise.VoiceSpeech to textAgents
• OpenAISpeech to speech with gpt-realtime, and gpt-4o-transcribe at about USD 0.006 a minute. Its Australian endpoint stores data here but processes it overseas.Speech to speechSpeech to textVoice
• Universal-Streaming speech to text at USD 0.15 an hour, in US and EU data zones only. No text to speech yet.Speech to text
• RimeVoices built for phone agents, Mist v3 at USD 0.03 per 1,000 characters, with one Australian voice, marlu. Self-hosting on Enterprise.Voice
• Fish AudioText to speech at USD 15 per million bytes of text and speech to text at USD 0.36 an hour. No region published.VoiceSpeech to text
• InworldTTS-2 at USD 25 per million characters and speech to text at USD 0.15 an hour. Residency in the EU and India on Enterprise.VoiceSpeech to text

Providers closing or closed

Hume is ending its text to speech and EVI voice APIs on 13 November 2026, and its docs say account data is deleted after that date. Do not start a new build on either.

PlayHT's API was switched off on 26 July 2025 after Meta bought the company, according to migration notices published by other voice vendors. Amazon's Nova Sonic speech to speech model is not offered in Sydney; the nearest Bedrock region for it is Tokyo.

Australian accent text to speech

These providers publish Australian English voices by name.

• Microsoft Azure: 15 en-AU neural voices, including Natasha and William.
• Amazon Polly: Olivia as a generative voice, and Nicole, Olivia and Russell as neural voices.
• Google Cloud: en-AU voices on Chirp 3 HD.
• Rime: one Australian voice, marlu.
• ElevenLabs: Australian voices in its voice library, plus cloning a voice from your own recording.
• Cartesia: cloning from about 10 seconds of audio on Pro and above.

Choosing a provider for an Australian voice agent

Decide first where the audio may be processed. If speech has to stay in Australia without an Enterprise contract, the shortlist is Deepgram, Azure, Amazon and Speechmatics.

Measure latency from the place your calls are answered. Advertised figures of 75 to 100 ms are measured at the vendor's servers; from Sydney we see 450 to 500 ms at the median.

• Read the concurrency limit on the plan you will buy. Cartesia Scale allows 15 voice streams and Deepgram's Flux TTS allows 5 on its Australian endpoint below Enterprise.
• Check what the residency setting leaves out: voice cloning, support access and billing records sit outside it at more than one vendor.
• Ask each vendor for its model retirement schedule. Cartesia retires sonic-turbo and sonic-2 after 20 October 2026.
• Price per minute of call, not per character. At 900 characters a spoken minute, the voices of the three main providers cost between USD 0.027 and USD 0.072 a minute.

The providers our voice agents run

Our voice agents run the language model in Sydney on AWS Bedrock. Speech recognition uses Cartesia Ink-2 by default and can move to Deepgram in Sydney one agent at a time. The default voice is ElevenLabs, with Cartesia as the fallback.

Our sub-processors page lists every provider and where each processes data, and our enterprise deployment pages show a build with the whole call in your own AWS account.