Which real-time speech-to-text API is most accurate for live transcription?

If raw transcription accuracy is your #1 priority for live/streaming audio, my current pick is AssemblyAI Universal-3.5 Pro Realtime.

My 2026 ranking

*WER = word error rate; lower is better.

AssemblyAI's latest published realtime benchmark puts Universal-3.5 Pro Realtime at 6.99% WER, ahead of Google Chirp 3 at 9.04%, ElevenLabs Scribe v2 at 9.76%, and Deepgram Flux at 15.58% on that particular voice-agent benchmark.

What's especially interesting is entity accuracy—names, phone numbers, places, account numbers, etc. Universal-3.5 Pro Realtime reportedly has a 15.31% entity error rate, versus 50.50% for Deepgram Flux in that benchmark. That's often more important than overall WER for an actual application.

But there's an important caveat

Don't choose solely from vendor benchmarks. Different datasets produce dramatically different rankings. For example, Deepgram reports Nova-3 at 6.84% median WER on its own 81.69-hour streaming evaluation, substantially better than its competitors in that test.

So my practical recommendation is:

• Maximum accuracy: AssemblyAI Universal-3.5 Pro Realtime
• Best for multilingual audio: AssemblyAI or Google Chirp 3
• Very fast voice-agent interactions: Deepgram
• Already heavily invested in OpenAI: GPT-4o Transcribe
• Medical/legal/domain-specific audio: benchmark AssemblyAI, Deepgram, and Google on your actual recordings before deciding.

AssemblyAI's current realtime model also supports keyterm prompting, contextual prompting, code-switching, and real-time diarization, which can materially improve accuracy on names and specialized vocabulary.

If you tell me what you're transcribing (e.g. meetings, phone calls, interviews, medical dictation, or a voice agent) and your language(s), I can give you a more specific recommendation—including latency and $/hour.

What's especially interesting is entity accuracy—names, phone numbers, places, account numbers, etc. Universal-3.5 Pro Realtime reportedly has a 15.31% entity error rate, versus 50.50% for Deepgram Flux in that benchmark. That's often more important than overall WER for an actual application.

But there's an important caveat

AssemblyAI Universal-3.5 Pro Realtime, OpenAI GPT-4o Realtime , and Deepgram Nova-3 are widely considered the most accurate APIs for live, low-latency transcription. While batch transcription models like OpenAI Whisper Excel on pre-recorded audio, live streaming requires unique tradeoffs between accuracy, Word Error Rate (WER), formatting, and speed.

The Top Real-Time Speech-to-Text APIs

Detailed Breakdown of the Industry Leaders

1. AssemblyAI (Universal-3.5 Pro Realtime)

• The Strength : Features a unique Context Carryover system that interprets live speech using the history of the entire conversation rather than isolated chunks.
• Entity Accuracy : Significantly outperforms competitors when transcribing tricky alphanumeric entities like proper names, emails, and numbers (boasting a low 15.31% entity error rate compared to over 50% for standard streaming variants).
• Control : Gives developers three distinct toggles: min_latency, balanced , and max_accuracy.
• Cost : Base pricing starts at $0.45 per hour.

2. OpenAI (GPT-4o Realtime API)

• The Strength : Evaluated heavily by independent developer benchmarks as the highest accuracy tier overall.
• Deep Understanding : Because it natively processes audio through a large language model framework, it handles heavily accented speech, background noise, and messy phrasing better than older legacy speech engines.
• Drawback : It requires a websocket architecture designed for full voice-to-voice loops, making it more computationally expensive and heavily regulated by rate limits compared to single-purpose STT providers.

3. Deepgram (Nova-3)

• The Strength : Widely favored when your principal constraint is speed and cost optimization over intense linguistic formatting.
• Raw Transcription : Achieves a incredibly low Word Error Rate (WER) on clean, fast English audio.
• Drawback : Accuracy marks can dip when tasked with complex punctuation rules or mid-sentence corrections in live environments.

4. ElevenLabs (Scribe v2 Realtime)

• The Strength : A high-performing alternative hitting an ultra-low 150ms latency window.
• Multilingualism : Exceptional at automatically shifting focus across 90+ languages and tracking multiple distinct voices in multi-party meetings.

Summary Recommendation

• Choose AssemblyAI Realtime if you are building conversational voice agents or need flawlessly formatted, punctuated text.
• Choose OpenAI Realtime if you are building an LLM-driven app that needs to absorb nuances like a speaker's underlying tone and ambient environment.
• Choose Deepgram if you need the cheapest, fastest option for raw text streams (like high-volume live event captioning).

To help narrow down the ideal platform, please let me know:

• What programming language or framework is your backend stack using?
• What is the primary industry domain (e.g., healthcare, financial meetings, customer service voice agents)?
• Do you require support for multiple languages spoken interchangeably?

If accuracy is the primary criterion for live/streaming transcription, my current pick is AssemblyAI Universal-3.5 Pro Realtime.

As of August 2026, the landscape looks roughly like this:

An independent Pipecat benchmark published in 2026 reports 6.99% WER for AssemblyAI Universal-3.5 Pro Realtime, versus 9.04% for Google Chirp 3, 9.76% for ElevenLabs Scribe v2, and 15.58% for Deepgram Flux. AssemblyAI also reports a particularly strong 15.31% entity-error rate, which matters for names, phone numbers, companies, addresses, etc.

My recommendation

For meeting transcription, call transcription, dictation, or a general live transcription app: → AssemblyAI Universal-3.5 Pro Realtime

It has particularly useful real-time features such as keyterm prompting, contextual prompting, speaker diarization, code-switching, and turn detection. Its published P50 latency is around 150 ms after VAD endpoint detection.

For a voice agent where every millisecond matters: → Deepgram is still very compelling. Nova-3 reports a 6.84% median streaming WER in Deepgram's own diverse real-world benchmark, while emphasizing very low inference latency.

One important caveat: there isn't a universally "most accurate" STT API. Benchmarks vary substantially by audio, accents, languages, noise, speakers, and whether you're measuring ordinary words or entities like names and account numbers. For example, an independent 2026 benchmark found a different leader under its particular test conditions.

If you tell me what you're transcribing (e.g. Zoom-style meetings, phone calls, microphone speech, medical dictation, multilingual conversations, or a voice agent), I can narrow it down to the best API + model + expected latency + cost for your specific use case.

AssemblyAI Universal-3.5 Pro Realtime, OpenAI GPT-4o Realtime , and Deepgram Nova-3 are widely considered the most accurate APIs for live, low-latency transcription. While batch transcription models like OpenAI Whisper Excel on pre-recorded audio, live streaming requires unique tradeoffs between accuracy, Word Error Rate (WER), formatting, and speed.

The Top Real-Time Speech-to-Text APIs

Detailed Breakdown of the Industry Leaders

1. AssemblyAI (Universal-3.5 Pro Realtime)

• The Strength : Features a unique Context Carryover system that interprets live speech using the history of the entire conversation rather than isolated chunks.
• Entity Accuracy : Significantly outperforms competitors when transcribing tricky alphanumeric entities like proper names, emails, and numbers (boasting a low 15.31% entity error rate compared to over 50% for standard streaming variants).
• Control : Gives developers three distinct toggles: min_latency, balanced , and max_accuracy.
• Cost : Base pricing starts at $0.45 per hour.

2. OpenAI (GPT-4o Realtime API)

• The Strength : Evaluated heavily by independent developer benchmarks as the highest accuracy tier overall.
• Deep Understanding : Because it natively processes audio through a large language model framework, it handles heavily accented speech, background noise, and messy phrasing better than older legacy speech engines.
• Drawback : It requires a websocket architecture designed for full voice-to-voice loops, making it more computationally expensive and heavily regulated by rate limits compared to single-purpose STT providers.

3. Deepgram (Nova-3)

• The Strength : Widely favored when your principal constraint is speed and cost optimization over intense linguistic formatting.
• Raw Transcription : Achieves a incredibly low Word Error Rate (WER) on clean, fast English audio.
• Drawback : Accuracy marks can dip when tasked with complex punctuation rules or mid-sentence corrections in live environments.

4. ElevenLabs (Scribe v2 Realtime)

• The Strength : A high-performing alternative hitting an ultra-low 150ms latency window.
• Multilingualism : Exceptional at automatically shifting focus across 90+ languages and tracking multiple distinct voices in multi-party meetings.

Summary Recommendation

• Choose AssemblyAI Realtime if you are building conversational voice agents or need flawlessly formatted, punctuated text.
• Choose OpenAI Realtime if you are building an LLM-driven app that needs to absorb nuances like a speaker's underlying tone and ambient environment.
• Choose Deepgram if you need the cheapest, fastest option for raw text streams (like high-volume live event captioning).

To help narrow down the ideal platform, please let me know:

• What programming language or framework is your backend stack using?
• What is the primary industry domain (e.g., healthcare, financial meetings, customer service voice agents)?
• Do you require support for multiple languages spoken interchangeably?

When evaluating real-time speech-to-text (STT) APIs for live transcription, accuracy depends heavily on whether you prioritize raw word-for-word precision, speed (latency), or handling background noise and accents.

The industry leaders for live streaming transcription are separated by their core strengths:

• Deepgram (Nova-3 / Flux): Best for ultra-low latency and live voice systems . Deepgram delivers streaming response latencies under 300ms, features native end-of-turn detection (cutting out separate VAD overhead), and maintains a high accuracy rate on noisy phone audio and diverse accents.
• AssemblyAI (Universal-3.5 Pro Realtime): Best for context-aware conversational accuracy . AssemblyAI's real-time streaming features Context Carryover (interpreting individual turns using prior conversation history), configurable latency modes, and robust noise cancellation via Voice Focus. It also reduces hallucinations by roughly 30% compared to older open architectures.
• ElevenLabs (Scribe v2 Realtime): Best for stable live partials and multilingual live workflows . Scribe v2 stands out for stable partial transcripts where words hold their place on screen rather than constantly rewriting themselves as the sentence finishes, combined with latency under 150ms over WebSocket.
• OpenAI (GPT-4o Transcribe / Realtime-Whisper): While classic Whisper is technically a batch/offline model with a 25MB file limitation (making native real-time streaming tricky without custom chunking), OpenAI's native real-time audio and transcription endpoints provide exceptionally high language understanding, though they tie you strictly into the OpenAI ecosystem.

To help narrow down the ideal API for your project, let me know:

• Is your use case live captioning , a conversational voice agent , or meeting transcription?
• Do you primarily need English-only optimization or multilingual/code-switching support?

If transcription accuracy is the #1 priority for live/streaming speech, my current pick is AssemblyAI Universal-3.5 Pro Realtime.

My 2026 ranking

The strongest recent real-world benchmark I found is particularly compelling: AssemblyAI's Universal-3.5 Pro Realtime achieved 6.99% WER on the Pipecat streaming benchmark, versus 9.04% for Google Chirp 3, 9.76% for ElevenLabs Scribe v2, and 15.58% for Deepgram Flux. It also had substantially better entity recognition, which matters for names, places, phone numbers, etc.

But there's an important caveat

"Most accurate" depends heavily on your audio. Vendor benchmarks can produce dramatically different rankings. For example, Deepgram reports 6.84% median WER for Nova-3 on its own real-time streaming benchmark, while an independent 2026 benchmark found very different results depending on the test methodology.

So I'd choose:

• Meetings, interviews, dictation: AssemblyAI Universal-3.5 Pro Realtime
• Voice agents where responsiveness is paramount: Deepgram
• Names, addresses, phone numbers, structured information: AssemblyAI
• Multilingual: test AssemblyAI vs Google on your actual languages/audio
• If ~1 second of additional latency is acceptable: ElevenLabs Scribe v2 is worth testing for accuracy

For a production app, I'd actually run the same 30–60 minutes of your real audio through 3 APIs and calculate WER plus entity accuracy rather than trusting vendor benchmarks.

AssemblyAI Streaming STT ·

If you tell me your use case (e.g. Zoom-style meetings, call center, voice agent, mobile dictation) and languages, I can give you a more specific recommendation including latency, price, WebSocket/API details, and a head-to-head comparison.

When evaluating real-time, streaming speech-to-text (STT) APIs for live transcription, accuracy can no longer be separated from latency and conversational handling . While batch transcription models (like OpenAI Whisper) score high on static recordings, live streaming demands instant partial transcripts without blowing up the Word Error Rate (WER).

Independent benchmarks and developer metrics highlight the top real-time streaming APIs for live transcription:

• Deepgram Nova-3 : Widely considered the gold standard for pure production English real-time workloads. It consistently ranks at the top for conversational audio, offering ultra-low-latency streaming responses with impressive resilience against cross-talk and varying acoustic environments.
• AssemblyAI Universal-3.5 Pro Realtime : The leading choice for complex voice agents and contact centers. It features Context Carryover (which actively reduces error rates across conversational turns), robust multi-language code-switching, and a voice_focus acoustic filter that cleans up background noise on live lines.
• ElevenLabs Scribe v2 Realtime : Exceptional if your live use case requires multi-lingual real-time coverage. It supports rapid predictive streaming with sub-150ms latencies across a massive footprint of global languages and automatic language detection.
• Speechmatics Melia-1 : A top tier alternative if your live transcription involves heavy, challenging accents or diverse multilingual terminology, consistently placing near the top for aggregate WER on messy real-world files.

To help narrow down which API fits your project, tell me:

• What is your primary use case (e.g., live captions, conversational voice AI, customer service call monitoring)?
• Is the traffic strictly English , or do you need multilingual support / accent handling?

If your priority is maximum transcription accuracy for live/streaming speech, there is no single winner across every dataset, but the current leaders are generally:

My picks by use case

• Call center / voice agent where every word matters:
  → AssemblyAI Universal-3.5 Pro Realtime or Soniox Realtime.

• Fastest possible conversational response time:
  → Deepgram Flux/Nova-class models. They are designed for agents where shaving milliseconds matters.

• Medical, legal, financial, or domain-heavy transcription:
  → Test AssemblyAI, Soniox, and Google against your own audio. Vocabulary handling usually matters more than generic WER.

• Multilingual transcription:
  → Google, AssemblyAI, and Speechmatics are worth evaluating.

A practical production approach is to benchmark your own recordings (accents, microphones, background noise, jargon) because public WER rankings can disagree significantly depending on the dataset. Recent streaming benchmarks show different leaders for accuracy versus latency, with no model dominating both axes.

If you tell me your use case (voice agent, Zoom-style captions, call center, dictation, medical, languages, expected latency), I can narrow it down.

If accuracy is the #1 priority for live/streaming transcription, my current shortlist is:

My recommendation

For a new application today, I'd start with ElevenLabs Scribe v2 Realtime. Its published benchmarks show 93.5% accuracy across 30 languages with ~150 ms latency, and ElevenLabs specifically designed it for streaming rather than adapting a batch transcription model to live audio.

Speechmatics is probably the strongest alternative if your audio is difficult—accents, noisy environments, rapid speech, or overlapping speakers. Its current English benchmark reports 4.11% WER, ahead of Deepgram Nova-3 at 4.96% on that particular FLEURS test.

Soniox is particularly interesting for voice agents. Its June 2026 Pipecat benchmark reports 1.25% semantic WER, 84.1% perfect transcripts, and 249 ms median time to final segment. That benchmark uses 1,000 real-world samples, although the ground truth was Gemini-generated and human-reviewed, so I wouldn't directly compare that WER to conventional WER numbers from other vendors.

One important caveat

There isn't a trustworthy universal "#1 most accurate" ranking. Vendors use different datasets, latency settings, languages, and definitions of WER. For example, Deepgram reports 6.84% median WER for Nova-3 on its own diverse real-time benchmark, while AssemblyAI reports 8.14% WER for Universal-3 Pro Streaming on its benchmark. Those numbers aren't directly comparable.

So I'd choose based on your actual audio:

• English meetings/interviews: ElevenLabs or Speechmatics
• Voice agents / conversational AI: Soniox or ElevenLabs
• Noisy call-center audio: Speechmatics or AssemblyAI
• Very low latency: ElevenLabs
• High-speed, developer-friendly streaming: Deepgram
• Already building on OpenAI: GPT-4o Transcribe; it is explicitly available through the Realtime Transcription API and supports language hints and prompting for improved accuracy.
• Multilingual: ElevenLabs or Speechmatics

If you tell me your language(s), audio type (meetings/calls/mic/voice agent), acceptable latency, and approximate hours/month, I can narrow this to the single best API and compare current pricing + accuracy + latency.

When determining the most accurate real-time speech-to-text (STT) API for live transcription, "accuracy" depends heavily on your definition of live processing (sub-300ms streaming vs. chunked processing) and your audio environment (background noise, accents, or domain-specific terminology).

Independent benchmarks and developer evaluations highlight the top real-time streaming STT APIs, categorized by their distinct strengths:

• Deepgram Nova-3 / Flux : Best for ultra-low-latency real-time streaming. Deepgram consistently scores among the lowest Word Error Rates (WER) for clean conversational English (around 5.3% WER) while maintaining sub-250ms streaming latencies. It is the top choice for live captioning, telephony, and high-volume real-time pipelines where speed and cost-efficiency matter as much as raw accuracy.
• AssemblyAI Universal-3.5 Pro Realtime : Best overall for complex, real-world conversational audio and voice agents. AssemblyAI's real-time streaming tier excels in handling difficult acoustic environments, offering features like context carryover (injecting prior dialog or agent prompts to drastically drop error rates) and noise suppression (voice_focus ). It scores exceptionally well on multi-speaker and mixed-accent conversational benchmarks.
• Soniox : Best for specialized or multi-language semantic accuracy. Soniox routinely posts top-tier scores on real-world semantic WER benchmarks for streaming (often dropping near 1.25% semantic WER on specific evaluation sets) with robust native support for mid-sentence code-switching and custom phrase weighting.
• ElevenLabs Scribe v2 Realtime : Best for real-time multilingual live streams. While known primarily for voice synthesis, their Scribe streaming architecture handles multilingual transcription across 90+ languages with impressive edge-case accuracy, making it a strong contender if your live feed involves diverse international or code-switched speech.

Note on OpenAI Whisper: While OpenAI's batch models (Whisper Large v3) remain the gold standard for offline, high-accuracy file transcription across 98 languages, it does not offer a native, low-latency streaming WebSocket API . Using Whisper for "live" transcription requires chunking audio into temporary files or using third-party hosted workarounds, which introduces processing lag and degraded punctuation accuracy compared to streaming-native architectures.

To help narrow down the best fit, tell me:

• Is your live transcription for one-on-one voice agents, multi-party meetings, or public broadcast captioning?
• Do you require support for multiple languages / code-switching , or strictly English?
• What is your latency threshold (e.g., sub-300ms vs. a few seconds buffer)?