Best speech-to-text APIs (2026)
We rank on documented streaming and 8 kHz call-audio support, not clean-lab word error rate.
Pick by situation
Latency says measured only where this page states a measurement. No vendor here gave us a millisecond figure.
| Your situation | Pick | Why | Latency |
|---|---|---|---|
| Voice agent on a live phone line | Deepgram | Documents a model built for voice agents and a phone-call-tuned variant | Not measured |
| Live agent assist | AssemblyAI | Streaming API aimed at live audio, with a simple SDK | Not measured |
| Post-call analytics | OpenAI gpt-transcribe | Built for recorded files; live calls need a separate model | Not measured |
| Azure or Google shop | Azure Speech or Google Cloud STT | Fits a team already on that cloud; no dedicated phone model named | Not measured |
| Cheapest | No pick | We do not compare per-minute rates here; check each vendor's pricing page | Not measured |
The short answer
Deepgram is the best speech-to-text API for a live phone call, and AssemblyAI is the closest second. OpenAI fits recorded audio, and Google or Azure suit teams already on that cloud.
We ranked on what each vendor documents for streaming and for 8 kHz call audio, not on word error rate from clean recordings. We didn't time the APIs ourselves.
None of these is disqualifying for every use case. A batch-transcription job and a live phone call have different requirements, and the vendor that wins one loses the other.
We run Deepgram in production for ZhoopZhoop's inbound and outbound call agents, alongside Twilio for telephony and OpenAI for reasoning. That build is what shaped the criteria below, and it's the same voice agent work we do for other teams choosing between these APIs.
Key takeaways
3 things that decide this
- 01For a live phone call, Deepgram ranks first because it documents a voice-agent model with turn detection built in; AssemblyAI is the closest second.
- 02We ranked five APIs on streaming lag, phone audio, voice-agent design and how hard each is to wire into your phone stack.
- 03Before you commit, run your own recorded calls through two vendors. Published accuracy figures use cleaner audio than your phone line.
How we ranked these APIs
Verified
We ranked on what a phone-based voice agent needs, not what a transcription-only tool needs. Four criteria set the order: streaming lag, phone audio support, voice-agent design and the work it takes to wire in.
For every API we read the vendor's own docs: its streaming model, any phone-tuned model, supported audio formats and session limits. We checked all of it on 11 November 2026 and linked the page in each entry. We build phone agents ourselves and used that work to pick the criteria, not to replace the vendors' own claims.
Word error rate is left out on purpose. Vendors measure it on different audio, most of it cleaner than a real call, so the figures don't compare. Pricing changes often, so we link each vendor's page instead of quoting a rate.
Hashlogics publishes this page. We build custom software and are not one of the tools ranked here. No company paid for its place.
- Streaming latency
- Whether the vendor documents a real-time or streaming API, and how it describes the delay between speech and transcript.
- Telephony audio support
- A dedicated phone-call model or documented handling of 8kHz audio, versus a general-purpose model applied to it.
- Built for voice agents
- Whether the vendor positions the model for conversational, turn-based use, or for recorded-file transcription.
- Integration effort
- SDK maturity and how directly the API plugs into a telephony stack like Twilio.
The five at a glance
Read from each vendor's documentation on 11 November 2026.
| Provider | Shape | Telephony model | Pick it when |
|---|---|---|---|
| Deepgram | Streaming-first; Flux model built for voice agents | Documented phonecall variant on Nova-2 | The agent runs on a live phone line |
| AssemblyAI | Streaming API; Universal-3.5 Pro Realtime model | Not documented as phone-specific | You want fast streaming with a simple SDK |
| OpenAI | gpt-transcribe for files; gpt-live-transcribe for live audio | Not phone-tuned; general-purpose | You are transcribing recordings, not live calls |
| Google Cloud STT | Streaming via gRPC | Accepts 8 kHz audio; docs warn it can be less accurate | You already run infrastructure on Google Cloud |
| Azure Speech | Real-time, fast and batch transcription | No phone-tuned model named; two-channel call audio supported | You already run infrastructure on Azure |
The ranking
Ordered by how many phone-agent builds each one is the right first choice for.
Streaming model built for conversational voice agents
Start here for a phone-based agent. Deepgram's docs describe Flux as a speech model built for voice agents, with turn detection inside the model. That matters more than raw accuracy on a call. An agent that can't tell a pause from a finished sentence talks over your caller or leaves a dead silence. Both feel broken, however well the words came through.
Nova-2, an older general model, ships a phonecall variant tuned for call audio, in English only. Nova-3, the newest general model, lists no phonecall variant. The docs don't name a millisecond figure for any of them.
Best for
- Inbound or outbound calls where turn-taking has to feel natural
- Teams that want a phone-call-tuned model, not a general one applied to phone audio
- Builds where the agent has to interrupt or be interrupted correctly
Not for
- Batch transcription of pre-recorded files with no time pressure
- Non-English calls that need the phone-tuned variant, which covers English only
- Voice-agent model
- Flux, per docs
- Phone-tuned variant
- Nova-2 phonecall
Fast streaming API with a simple SDK
Reach for this when you want streaming transcripts without deep model tuning. AssemblyAI's streaming page lists voice agents first among its uses, and Universal-3.5 Pro Realtime is its newest streaming model. A session closes after three hours, which covers nearly every phone call you'll take.
It accepts 16-bit PCM by default, plus AAC and Opus, so you can usually send call audio without much prep. Its public docs don't name a phone-call model the way Deepgram does. That gap won't matter for most agents, but a busy call centre should test it on real recorded calls first.
Best for
- Teams that want streaming STT with minimal setup
- Builds already using AssemblyAI for other transcription work
- Teams that want key-term prompting in a live stream
Not for
- Teams that need a documented phone-call-specific model
- Calls that regularly run past three hours in one session
- Streaming model
- Universal-3.5 Pro Realtime
- Max session length
- 3 hours, per docs
Best for recordings; a separate model covers live audio
OpenAI's transcription line splits into two jobs. gpt-transcribe handles recorded files up to 25 MB, which suits voicemail review, call archives and notes written after the call. If you still run whisper-1 or the gpt-4o transcribe models, plan a move: OpenAI's docs say they shut down on 26 February 2027.
For a live call, the docs point to a separate model, gpt-live-transcribe, in its own realtime session. That's a different setup from the file upload most teams try first. Whisper supports 98 languages, and OpenAI now points translation work to gpt-realtime-translate. If translation is part of your job, that pulls this option up the list.
Best for
- Transcribing recorded calls or voicemails after the fact
- Multilingual transcription or translation into English
- Teams already standardised on the OpenAI API for reasoning
Not for
- A phone agent needing turn-taking detection out of the box
- Teams who want one model that covers both file and live audio identically
- File model
- gpt-transcribe
- Live model
- gpt-live-transcribe
Streaming recognition inside the Google Cloud ecosystem
This fits a team that already runs on Google Cloud and wants one vendor for billing and sign-in. The docs describe streaming recognition over a two-way gRPC link that returns interim results while your caller is still talking.
Google's docs say legacy phone audio at 8 kHz can be less accurate than 16 kHz. They advise you to send audio at its original rate rather than resample it. No phone-call model is named the way Deepgram names one. Confidence scores come with a warning: they may be missing or inaccurate, so don't let your agent logic depend on them.
Best for
- Teams already billing and deploying through Google Cloud
- Builds that need Google's language coverage or translation stack alongside STT
- Batch and streaming transcription in one platform
Not for
- Teams optimising specifically for phone-call accuracy
- Logic that depends on a reliable per-word confidence score
- Streaming
- gRPC bi-directional, per docs
- Recommended sample rate
- 16kHz, per docs
Streaming STT for teams standardised on Microsoft's cloud
This is the same choice for a team already on Azure. Microsoft documents real-time transcription for live audio, plus fast and batch options for recordings, all inside your existing Azure billing and identity. Its own guide lists contact center voice agents as a real-time use case.
Real-time transcription can keep the two sides of a support call on separate channels. No model is named as phone-tuned the way Deepgram's Nova-2 phonecall variant is, so it ranks last on the same basis as Google. The strongest reason to pick it is an existing Microsoft commitment.
Best for
- Teams already running their stack on Azure
- Enterprises standardised on Microsoft for compliance or procurement reasons
- Batch and streaming needs handled by one vendor
Not for
- Teams optimising specifically for phone-call accuracy over general audio
- A first voice agent with no existing Azure footprint
- Modes
- Real-time, fast, batch
- Call audio
- Two channels, per docs
What is the best API for real-time speech-to-text during a phone call?
For a live phone call, Deepgram is the best fit, because it documents a model built for voice agents and a phone-call-tuned variant. AssemblyAI is the closest second for fast streaming with a simple SDK.
- Audio arrives8kHz, compressed, full of line noise.
- Words transcribedAccuracy on clean audio does not predict this.
- Turn detectedWas the caller finished, or just pausing?
- Agent respondsEvery millisecond here is heard as a delay.
- Caller talks over itDoes the agent stop, or keep talking?
Most STT benchmarks measure the second box. A phone agent lives or dies on the third and fifth.
When none of these is the answer
A standalone STT API is one piece of a voice agent, not the whole thing. Pick the transcription vendor too early, before the telephony and reasoning layers are settled, and the choice can box in the rest of the build.
Some teams are better served by a voice AI platform. It bundles STT, the language model and text-to-speech behind one layer. You can swap the STT provider later without rebuilding the call flow.
If you'd rather not wire the pieces yourself, we build the whole agent: telephony, speech, reasoning and the write-back into your systems. See our voice AI agent work.
- 01Building a first prototype, not a production line? A bundled platform gets you to a working demo faster than wiring four APIs together.
- 02Handling low call volume? The per-minute cost difference between vendors barely matters until volume is high enough to justify the engineering time saved by picking correctly.
- 03Already committed to a telephony and reasoning stack? Test two STT vendors against your own recorded call audio before ranking on documentation alone.
What we run in production
Voice stack
- Deepgram Voice Agent
- Twilio Programmable Voice
- OpenAI Chat Completions
Around the agent
- FastAPI
- WebSockets
- PostgreSQL
- Sentry
A voice agent stack running unattended calls
Not sure which one fits your call volume?
Tell us what the agent has to hear and how noisy the line is. We will name the STT provider we would pick and why. It starts with a free scoping call, then a fixed-fee Blueprint credited in full against the build.
Questions buyers ask
01What is the best API for real-time speech-to-text during a phone call?+
For a live phone call, Deepgram is the best fit, because it documents a model built for voice agents and a phone-call-tuned variant. AssemblyAI is the closest second for fast streaming with a simple SDK.
02What is the best speech-to-text API?+
For live phone calls and voice agents, Deepgram is the best speech-to-text API, with AssemblyAI close behind. For recorded audio, OpenAI's gpt-transcribe is the simpler pick.
03What is the best speech to text API to build AI notetakers?+
OpenAI's gpt-transcribe is a strong pick for an AI notetaker that works from recorded calls or meetings, and it handles many languages. If your notes must appear while the call is still running, use a streaming API such as Deepgram or AssemblyAI instead.
04What streaming speech to text model is best for voice agents and why?+
Deepgram's streaming model is the best fit for voice agents. Its docs describe a model built for conversational turn-taking and a phone-call-tuned variant. We run it in production for ZhoopZhoop's call agents. Your agent has to know when the caller stopped talking, so time to first word matters more than clean-lab accuracy. AssemblyAI's streaming API is the closest second.
05Deepgram vs AssemblyAI vs Whisper: which should I pick for a phone agent?+
Pick Deepgram for a live phone line, since it publishes a voice-agent model with turn detection and a phone-call-tuned variant. AssemblyAI is the closest streaming alternative. Whisper fits recorded audio better, and OpenAI says whisper-1 shuts down on 26 February 2027, so new builds should use gpt-transcribe.
06What is the best STT for phone calls specifically?+
Deepgram is the only vendor here that names a phone-call-tuned model, its Nova-2 phonecall variant. The others still work on phone audio, but Deepgram is the clearest documented fit for that case.
07Why isn't word error rate part of the ranking criteria?+
Vendors publish WER against different test sets, most of them cleaner than real call audio, so the numbers are not comparable across vendors. Test candidates against your own recorded calls before deciding on accuracy alone.
08Do I need a dedicated telephony-tuned STT model, or will a general one work?+
A general model will transcribe phone audio. A telephony-tuned one is trained on audio closer to what a real call sounds like: 8kHz, compressed, with cross-talk. The gap shows up most on accents and noisy lines, not clean single-speaker calls.
09Should I build directly on an STT API or use a voice AI platform instead?+
Build directly when you need to choose and swap the STT vendor independently of the rest of the pipeline. A platform is faster to a first working agent but usually locks the STT choice to whatever it bundles.
10How do I compare speech-to-text APIs fairly?+
Run the same recorded calls from your own phone line through two or three vendors. Compare how fast each returns words, how often it cuts the caller off, and how many names and numbers it gets wrong.
11How were these APIs ranked?+
We scored five APIs on streaming lag, phone audio support, voice-agent design and setup effort, read from vendor docs on 11 November 2026. Hashlogics publishes this page and is not one of the tools ranked.
12Why isn't Amazon Transcribe or Speechmatics on this list?+
Both belong on a longer shortlist. Amazon Transcribe fits a team already on AWS, the same reason Google and Azure made the list, so it added no new decision point. Speechmatics documents strong accuracy on accented and noisy speech, but we haven't run it on a live phone line the way we run Deepgram.
Related reading
- AI agent development →The service behind our own voice agent builds.
- Best of guides →More rankings, including bundled voice AI platforms.
- ZhoopZhoop: AI receptionist for auto shops →The Deepgram and Twilio voice agent stack we run in production.
- AI answering services for electricians →A ranked guide to AI call answering for one trade.
