8 Speech-to-Text APIs for AI Voice Agents Compared

The speech-to-text API is the ears of an AI voice agent. For live calls you need streaming transcription, fast final results, reliable end-of-turn signals and good accuracy on phone audio.

The eight APIs compared here are AssemblyAI, Deepgram, Soniox, Speechmatics, Azure AI Speech, Amazon Transcribe, Google Cloud Speech-to-Text and OpenAI's realtime transcription.

Published benchmarks help shortlist, but the right choice depends on your callers, languages and audio. Test on your own recordings before deciding.

How We Evaluated These APIs

We read each provider's current documentation on November 2, 2026, and the public Pipecat speech-to-text benchmark published on GitHub. We did not run our own tests. Each API was compared on the same criteria:

• Streaming support: real-time transcription over a live connection

• Turn handling: whether the API signals when a caller has finished speaking

• Deployment: cloud regions, private or on-premise options where documented

• Published results: word error rate (WER) and median time to final transcript from the Pipecat benchmark, where listed

• Fit: the call types where each one is a sensible candidate

Benchmark figures come from Pipecat's 1,000-sample English test set. They are third-party results, not ours, and may not reflect 8 kHz phone audio, accents or noisy lines. The order is not a ranking.

Published Benchmark Snapshot

Source: Pipecat stt-benchmark README, read November 2, 2026. Results change as models update.

1. AssemblyAI Universal Streaming

AssemblyAI documents a streaming speech-to-text API for real-time applications, including voice agents, with end-of-turn detection.

• Fits: English-heavy support and booking calls where accuracy on names, numbers and emails matters

• Check: session billing rules and language coverage for your callers

• Docs: AssemblyAI Universal Streaming

2. Deepgram (Nova-3 and Flux)

Deepgram's Flux documentation describes a conversational model with built-in end-of-turn detection for voice agents. Nova-3 is its general speech model.

• Fits: latency-sensitive agents where fast turn-taking matters most

• Check: accuracy on your domain terms, using keyterm features if needed

• Docs: Deepgram Flux quickstart

3. Soniox

Soniox documents real-time and async speech-to-text APIs, with support for more than 60 languages, plus translation and text-to-speech.

• Fits: multilingual lines where callers switch languages

• Check: data residency options for your region

• Docs: Soniox documentation

4. Speechmatics

Speechmatics documents real-time and batch transcription, an Agent STT mode for turn-based voice agents, and deployment as containers, a virtual appliance or on Kubernetes.

• Fits: regulated teams that need private or on-premise deployment

• Check: latency in your own deployment, not only the hosted service

• Docs: Speechmatics documentation

5. Azure AI Speech

Azure AI Speech documents real-time recognition from microphones, files and streams through its Speech SDK, with wide language support.

• Fits: teams already on Azure that want speech inside the same tenant and identity setup

• Check: the published median latency above is higher than the streaming specialists; test it on your call flow

• Docs: Azure AI Speech, recognize speech

6. Amazon Transcribe Streaming

Amazon Transcribe documents streaming transcription over HTTP/2 or WebSocket, for live audio such as calls.

• Fits: AWS-based contact center stacks and Amazon Connect environments

• Check: latency against your turn-taking budget, and partial-result settings

• Docs: Amazon Transcribe streaming

7. Google Cloud Speech-to-Text

Google Cloud documents streaming recognition for live audio through its Speech-to-Text API. Pipecat's benchmark lists two Google models with different results, so check which model you would use.

• Fits: Google Cloud stacks and multilingual deployments

• Check: model choice and regional endpoints

• Docs: Google Cloud streaming recognition

8. OpenAI Realtime Transcription

OpenAI documents a realtime transcription mode over its Realtime API, which streams audio and returns transcripts as speech arrives.

• Fits: teams already using OpenAI models for the agent's reasoning layer

• Check: latency and turn detection settings; published median latency is higher than the fastest options

• Docs: OpenAI realtime transcription

Our own product: Voxentis is LuMay's AI voice agent product, so it is not ranked here.

LuMay's voice agent pages list Deepgram, Azure Speech, Google STT, Whisper and AssemblyAI, and say transcription engines can be swapped. Check the current list on the product page.

How to Test STT on Your Own Calls

Stage 1: Collect audio. Gather 200 to 300 consented, real call segments covering your accents, noise levels and phone lines.

Stage 2: Label truth. Have people transcribe them, marking names, numbers, addresses and product terms.

Stage 3: Measure accuracy. Calculate WER overall and, separately, error on the entities your agent must capture.

Stage 4: Measure timing. Record time from end of speech to final transcript at the median and the 95th percentile. Tail latency causes awkward pauses.

Stage 5: Test turn-taking. Check whether the API cuts callers off while they spell an email or read a card number.

Stage 6: Check operations. Confirm regions, data retention, concurrency limits and pricing for your expected call volume.

Sample sizes are illustrative. See our guide to testing voice agent interruptions, latency and handoffs for the full conversation test.

Worked Example: Shortlisting for a Utility Hotline

Illustrative example. A utility's outage line handles English and Spanish callers, many on mobile phones outdoors.

• Must have: streaming with end-of-turn signals, Spanish support, US data processing

• Shortlist: three APIs whose documentation meets those needs

• Test: 250 labeled calls, scored on account-number accuracy and 95th-percentile latency

• Decision rule: pick the API with the best account-number accuracy among those under the team's latency limit

The winner on a public benchmark may not win here. Outdoor noise and account numbers are not what most benchmarks test.

Common Mistakes to Avoid

• Choosing on overall WER alone: errors on names and numbers matter more than filler words

• Testing studio audio: phone audio is narrower and noisier

• Ignoring tail latency: one slow turn in twenty still annoys callers

• Locking in one engine: prompts and parsers tied to one vendor's output make switching costly

• Skipping consent and retention review: call audio is often sensitive data

Limitations

Benchmark numbers in this article are third-party, from one English dataset, and will change as providers update models.

We did not verify pricing, which changes often. Use this comparison to build a shortlist, then decide on your own test results. For the wider stack, see our voice agent stack guide.

Next Step

If you want a voice agent without managing speech vendors yourself, review Voxentis by LuMay and its supported engines and integrations.

• AssemblyAI Universal Streaming

• Deepgram Flux quickstart

• Soniox documentation

• Speechmatics documentation

• Azure AI Speech, recognize speech

• Amazon Transcribe streaming

• Google Cloud streaming recognition

• OpenAI realtime transcription