Speech recognition has always been a crowded space. But in the last few years, the models have gotten faster, cheaper, and smarter. New architectures have entered the picture. And the baseline expectation for what "good enough" looks like has shifted dramatically.
If you're building a voice product in 2026, an AI meeting assistant, a contact center solution, a voice agent, anything that turns spoken words into structured data, you're probably starting your evaluation with the three names that have dominated this category: OpenAI Whisper, Google Cloud Speech-to-Text, and Amazon Transcribe. They're the bigger providers out there. The ones your infrastructure team probably already has an SSO login for.
We walk through each pairwise matchup, OpenAI Whisper vs Google STT, OpenAI Whisper vs Amazon Transcribe, and Google vs Amazon, covering accuracy, features, pricing, developer experience, and the real-world limitations that benchmark tables won't show you.
The state of ASR in 2026
Speech recognition as a category has split in two. On one side, basic transcription has become a commodity. Word Error Rates (WER) on clean, single-speaker English audio are now so low across the major providers that they've essentially stopped being a meaningful differentiator. If you're transcribing podcast episodes or pre-recorded interviews in a controlled environment, almost any modern API will do the job adequately. The race to the bottom on clean-audio WER is largely over.
On the other side, conversational audio remains genuinely hard, and the gap between providers here is wide. Multi-speaker calls, overlapping speech, accented voices in noisy environments, code-switching, speaker diarization, these problems haven't been solved by throwing more parameters at LibriSpeech benchmarks. They require different architectural choices, different training data, and different product decisions, and not every provider has made them.
This is the context that matters when evaluating OpenAI Whisper, Google STT, and Amazon Transcribe in 2026. Each has made meaningful moves. OpenAI shifted its transcription stack onto the GPT-4o architecture with gpt-4o-transcribe, launched a Realtime API now in general availability, and introduced a dedicated gpt-4o-transcribe-diarize endpoint. Google progressed through Chirp 2 to Chirp 3, adding a built-in denoiser and full streaming support, plus native diarization for batch jobs, closing much of the gap for production use cases beyond straight batch transcription. Amazon has been the quietest of the three on model innovation, but has continued deepening its lead on domain-specific tooling: call analytics, medical transcription, and AWS-native compliance workflows.
The practical result is that the Big Three are more capable than they were but capable in different ways, and still optimized around different assumptions about what audio looks like. Understanding those assumptions is the point of this guide. A few specialized providers, Gladia, Deepgram, AssemblyAI, ElevenLabs, have also made substantial moves and are increasingly competitive in specific use cases. We offer Solaria-1, built for full language breadth, code-switching, and real-time streaming, and Solaria-3, our newest model for real-world European business and contact-center audio, which currently ranks #1 on Switchboard, the hardest of its benchmark datasets, ahead of AssemblyAI, Deepgram, ElevenLabs, and others.
Accuracy and speed: where Whisper, Google STT, and Amazon Transcribe actually stand
The honest answer in 2026 is that WER numbers on public benchmarks are not the only buying signal. The gap between the best and worst performers on LibriSpeech or FLEURS has compressed to the point where it rarely drives purchase decisions.
What drives decisions (or should) is performance on your audio: accented speech, multi-speaker conversations, noisy environments, code-switching, etc. These are the variables that separate providers in production.
That said, here's where the independent benchmarks roughly land:
| Provider |
Model |
Approx. WER (independent benchmark*) |
Notes |
| OpenAI |
gpt-4o-transcribe |
~4.0% |
Clean-dataset number looks strong, real-world testing puts it mid-pack |
| OpenAI |
Whisper Large V3 |
~4.1 |
Open-source, still widely used for self-hosted deployments |
| Google |
Chirp 3 |
~4.3% |
Up from the original Chirp's 31.2% |
| Amazon |
Transcribe (standard) |
~4.1% |
Roughly tied with gpt-4o-transcribe, not behind it |
*Independent benchmark: Artificial Analysis, a blended real-world index (agent calls, conversational audio, earnings-call recordings). Figures current as of September 2026 and will move as the leaderboard updates. Check the live page before citing an exact number.
- OpenAI's own benchmark claims 2.46% WER for gpt-4o-transcribe on the FLEURS English subset, a clean, curated dataset. On Artificial Analysis's independent real-world benchmark, gpt-4o-transcribe lands at 4.0% WER, roughly mid-pack rather than a clear leader. The older Whisper Large V3 sits at 15.95% WER on the AMI meeting-audio dataset (real, overlapping, multi-speaker conversation), a much harder test than the clean benchmarks that made Whisper V3 look competitive.
- Google's Chirp 3 has closed the gap considerably from the original Chirp model. On Artificial Analysis's independent benchmark, Chirp 3 scores 4.3% WER against the original Chirp's 31.2%. Chirp 3 also adds a built-in denoiser for better handling of noisy audio. On language breadth, Chirp 3 supports 37 languages in general availability and 83 more in preview, around 120 total, still the widest coverage of the three.
- On Artificial Analysis's independent benchmark, Amazon Transcribe lands at 4.1% WER, essentially tied with gpt-4o-transcribe rather than behind it. Where Transcribe differentiates is in specialized verticals: Transcribe Call Analytics and Transcribe Medical support custom vocabulary and domain-specific language models, which AWS states can improve accuracy on domain terminology by up to 10-15% in some call-center scenarios. That's a features-driven improvement, not an independently verified vertical-vs-general-purpose benchmark.
The bottom line on accuracy: independent benchmarks now put OpenAI, Google, and Amazon within about half a point of WER on blended real-world audio. None is a clear leader on pure transcription accuracy anymore. On that same independent benchmark, our Solaria-3 scores 3.2% WER, ahead of gpt-4o-transcribe (4.0%), Amazon Transcribe (4.1%), and Chirp 3 (4.3%). The decisions that matter more are: does this provider handle speaker overlap? Does it do reliable diarization? What happens on accented speech that isn't well-represented in training data? Those are the tests you need to run with your own audio.
Beyond transcription: features that matter
The features that separate platforms for production use cases are the ones that come after the words appear on screen.
| Feature |
OpenAI |
Google Chirp 3 |
Amazon Transcribe |
| Real-time streaming |
Realtime API (GA Aug 2025) |
Full support in Chirp 3 |
Available |
| Speaker diarization |
gpt-4o-transcribe-diarize (dedicated endpoint) |
Native in Chirp 3 |
Up to 10 speakers |
| PII redaction |
Yes |
Yes |
Native |
| Custom vocabulary |
Via prompting |
Speech adaptation |
Custom language models |
| Medical transcription |
No |
No |
HIPAA-eligible, $0.075/min |
| Call analytics |
No |
No |
Transcribe Call Analytics |
| Built-in audio denoiser |
No |
Chirp 3 |
Noise reduction mode |
| Language support |
99 languages |
85+ languages (Chirp 3) |
100+ languages |
| Translation |
English output |
Chirp 2+ |
No |
| Zero data retention |
Optional |
Processes in memory |
With KMS encryption |
Real-time streaming is now available across all three providers, but with meaningful differences in implementation quality. Google's Chirp 3 added streaming as a first-class feature. OpenAI's Realtime API supports voice-to-voice interaction beyond pure transcription. Amazon has offered streaming for some time. Latency and reliability vary, if you're building real-time meeting assistants or voice agents, our guide to real-time transcription architecture walks through the infrastructure decisions that actually matter, including how sub-300ms latency is achieved in production.
Speaker diarization, identifying who said what, remains one of the most technically demanding capabilities in conversational speech, and one of the most underestimated when evaluating ASR providers. Understanding how diarization error rate (DER) is measured and what benchmark results actually mean in real environments is essential before making a decision here. Google Chirp 3 includes diarization natively. OpenAI has introduced gpt-4o-transcribe-diarize as a dedicated endpoint, though it requires audio chunking for files longer than 30 seconds. Amazon Transcribe's 10-speaker limit can constrain meeting transcription use cases with larger groups.
PII redaction and compliance is table stakes for enterprise contact center and healthcare use cases. All three providers offer it. Amazon Transcribe includes content redaction natively and has HIPAA-eligible medical transcription (priced separately at $0.075/minute). Google and OpenAI both support data processing agreements under GDPR and CCPA frameworks.
Code-switching, handling audio where speakers switch languages mid-sentence, is where general-purpose models still struggle. This is particularly acute in contact center environments and global enterprise products where multilingual teams are the norm, not the exception.
Custom vocabulary and model adaptation is available across all three. Google's Chirp 3 supports speech adaptation for domain-specific vocabulary. Amazon allows custom language models for domain-specific use. OpenAI allows prompting to steer transcription toward specific terminology. For a practical walkthrough of how to configure accuracy tuning, code-switching, and language detection settings in a production integration, see our getting-started guide.
Pricing: Whisper vs Google STT vs Amazon Transcribe
Pricing has remained relatively stable at the top level, but with more tiering at scale.
| Provider |
Model |
Price per minute |
Notes |
| OpenAI |
Whisper (hosted API) |
$0.006 |
Lowest commercial rate, 25MB file limit |
| OpenAI |
gpt-4o-transcribe / mini |
$0.006 / $0.003 |
Updated since launch, mini recommended for most use cases |
| Google |
Standard / Chirp (V2) |
$0.016 |
Chirp included at standard rate, +40% if data logging opt-out |
| Google |
Enhanced models |
$0.036 |
V1 API only |
| Amazon |
Tier 1 (standard) |
$0.024 |
Drops to $0.0078/min at 5M+ minutes/month |
| Amazon |
Medical transcription |
$0.075 |
HIPAA-eligible. domain-specific models |
OpenAI Whisper API remains the lowest commercial per-minute rate at $0.006/minute and $0.003/minute for gpt-4o-mini-transcribe. If you're seriously weighing build vs. buy, our decision framework for open-source vs. API walks through the full TCO calculation.
Google Cloud Speech-to-Text charges $0.016/minute for standard models, with Chirp included at that rate in V2. Volume discounts are available at scale but require contacting Google Cloud sales for specifics. New accounts get $300 in free credits, and 60 minutes per month remain free regardless. Worth noting: GCP's supporting infrastructure costs, storage, egress, IAM, Functions, can meaningfully add to the effective rate for teams not already on Google Cloud.
Amazon Transcribe uses a tiered structure: $0.024/minute at Tier 1, dropping as low as $0.0078/minute at 5M+ minutes monthly. Medical transcription is $0.075/minute. If you're already on AWS, the integration with existing billing and infrastructure can offset the higher sticker price.
Developer experience
OpenAI's API remains the easiest onboarding in the category. Less than six lines of code for a basic transcription. Clear documentation. No upfront credit card requirement for exploration. The tradeoff: the API is less mature for enterprise workflows, no S3 integration, limited webhook support, and the 25MB file size limit on the hosted API creates friction for longer audio. For a detailed technical comparison of how the OpenAI transcription API performs against a purpose-built provider in real production scenarios, our Whisper API vs. Gladia breakdown is worth reading before you commit to an integration.
Google Cloud Speech-to-Text is powerful and well-documented, but the onboarding process remains genuinely complex for developers not already embedded in GCP. The V2 API with Chirp 3 is cleaner than legacy V1, but you're still navigating IAM permissions, GCS bucket configurations, and regional availability constraints before you get your first transcript. For teams already on GCP, this overhead is invisible; for everyone else, it's real.
Amazon Transcribe follows the same pattern as Google: deeply capable, but the AWS onboarding experience is its own category of friction. The S3-upload, job-creation, poll-for-output pattern is fine if your data already lives on AWS. Starting from scratch, it's the slowest time-to-first-transcript of the three.
Privacy and security
All three providers have mature approaches to data security, and all three are appropriate for enterprise use cases with proper configuration.
- Amazon Transcribe uses TLS 1.2 for data in transit. Output is encrypted at rest by default with an AWS-managed S3 key, with customer-supplied AWS KMS keys supported as an option. Input media encryption is supported in the source S3 bucket. Both Amazon Transcribe and Transcribe Medical are HIPAA-eligible services, which requires a signed AWS Business Associate Addendum (BAA), not automatic compliance out of the box.
- For streaming and synchronous requests, Google processes audio in memory without persistent storage. For asynchronous (batch) requests, the resulting transcript, not the audio itself, is stored for about 5 days for retrieval. Data logging to improve Google's models is opt-in only. HIPAA coverage is available with a signed BAA, and SOC 2 is independently audited per product. Data residency controls, restricting processing to a single region, are a V2-specific capability. The legacy V1 API processes audio globally by default.
- OpenAI's own data controls documentation currently lists no default abuse-monitoring retention for the transcription endpoint, unlike some of OpenAI's other API endpoints, which do carry a 30-day window. Zero Data Retention is available and explicitly supported for transcription. Enterprise DPA agreements are available. Retention policies can change, so it's worth reconfirming the current setting directly against OpenAI's data controls page before processing sensitive audio.
Which one should you actually use?
There's no universal answer but some patterns can help you make an informed decision:
| Use case |
Recommended starting point |
Why |
| Highest accuracy on clean audio |
OpenAI gpt-4o-transcribe |
Leads on OpenAI's own FLEURS benchmark (2.46% WER, English); closer to mid-pack on independent real-world testing |
| Broadest language coverage + GCP already in stack |
Google Chirp 3 |
85+ languages, streaming, diarization, built-in denoiser |
| Call center / contact center analytics |
Amazon Transcribe Call Analytics |
Domain-specific models built for this; deep AWS integration |
| Medical transcription |
Amazon Transcribe Medical |
HIPAA-eligible; specialized vocabulary |
| Lowest cost at high volume |
Whisper (self-hosted) or Amazon Transcribe (Tier 3) |
$0 marginal cost vs $0.0078/min at scale, but model the TCO carefully |
| Voice agents / real-time streaming |
OpenAI Realtime API or Google Chirp 3 |
Both now GA for streaming; test latency with your use case |
| Conversational / multi-speaker audio |
Test all three on your own audio, or see how purpose-built providers compare |
Benchmarks don't predict real-world conversational performance |
If you're already on GCP or need the broadest language coverage with enterprise-grade compliance infrastructure, Google's Chirp 3 is a significant step up from where Google's ASR stood two years ago. It's competitive on accuracy, has real streaming support now, and the diarization improvements are meaningful.
If you're on AWS, have call center or medical transcription use cases, or need to process audio at very high volume with predictable pricing, Amazon Transcribe's ecosystem integration and specialized models justify the evaluation time.
If your use case involves conversational speech, multi-speaker audio, overlapping speech, real-world noise, code-switching, or meeting intelligence, this is where it might be worth extending your search beyond the Big Three. Their general-purpose architectures weren't designed around the problems that conversational audio creates. It's worth looking at providers purpose-built for this. Our Solaria-1 model was specifically designed for multi-speaker, multilingual, real-world audio with native code-switching across 100+ languages. Solaria-3, our most accurate model for European business and contact-center audio, currently ranks #1 on Switchboard, the hardest of its benchmark datasets, ahead of AssemblyAI, Deepgram, ElevenLabs, and others. Our blind STT comparison tool is a fun and useful way to test your own audio. It strips out provider branding so you pick the better transcript before seeing who produced it, a useful gut check before you commit engineering time.
The benchmark problem no one talks about
Published WER numbers are designed to look good, not to tell you what will happen in your production environment. Every major ASR provider curates the benchmark datasets they publish against. The audio is often clean, the speakers are often native, the recording conditions are controlled. Real-world audio doesn't look like that.
The only evaluation that matters for your use case is the one you run on audio that actually represents what your users will throw at the system: multi-speaker calls with background noise, accented speech that doesn't appear in training data, technical vocabulary, etc. We open-sourced our benchmarking methodology across 8 providers and 74 hours of real audio for exactly this reason, so you can see what performance looks like on production-representative data, and reproduce every result yourself.
Start with €50 in free credits and have your integration in production in less than a day. Test Gladia on your own multilingual audio to see how it handles language detection, accent-heavy speech, and code-switching.
FAQs
What is the difference between OpenAI Whisper, Google Speech-to-Text, and Amazon Transcribe?
OpenAI leads on raw accuracy for clean audio, Google leads on language breadth, and Amazon leads on domain-specific tooling. OpenAI's hosted API (gpt-4o-transcribe) posts the lowest WER on clean single-speaker audio and the simplest onboarding. Google Chirp 3 covers 125+ languages with native diarization, a built-in denoiser, and streaming, but onboarding is complex outside GCP. Amazon Transcribe is mid-tier on general accuracy but strongest for call center and medical use cases through its specialized models and deep AWS integration. For conversational, multi-speaker, or multilingual audio, none of the three was designed around that problem specifically, which is where purpose-built providers become worth testing.
How does Google Chirp 3 compare to OpenAI's models?
Chirp 2 benchmarks at around 11.6% WER, better than Whisper Large V3 on standard tests, though behind gpt-4o-transcribe. Chirp 3 improves further and adds a built-in denoiser, native diarization, and full streaming support. Google's main advantages are language breadth (125+ languages) and deep GCP ecosystem integration. Its main friction point is onboarding complexity for teams not already on Google Cloud.
How do I run a fair evaluation of ASR providers for my use case?
Before committing engineering time to a full integration, it's worth grounding your evaluation in something more representative. We tested 8 providers across 74 hours of real production audio and open-sourced the entire methodology so every result can be independently reproduced. You can run your own test, measure WER on multi-speaker segments specifically, check diarization accuracy under real noise, and pay attention to the accents and languages your users actually speak.
Key terms glossary
Word error rate (WER): the percentage of words a transcript gets wrong (substituted, deleted, or inserted) against a verified reference transcript. A single WER number only means something alongside the audio condition it was measured on, since clean single-speaker audio and noisy multi-speaker calls produce very different numbers from the same model. What is WER
Diarization error rate (DER): the standard metric for how accurately a system attributes speech to the right speaker across a multi-speaker recording. Lower DER means fewer mixed-up turns and mislabeled speakers.
Diarization: the process of identifying who spoke when in a recording, distinct from transcription itself. It's typically run as a post-processing step on the full recording rather than live. What is diarization
Code-switching: when a speaker changes language mid-conversation or mid-sentence, common in bilingual teams and global contact centers. Models trained mostly on single-language audio often fail silently here instead of flagging the switch. What is code-switching
PII redaction: detecting and masking personally identifiable information (names, phone numbers, card numbers) in a transcript when the feature is turned on. It's a configurable option rather than a default behavior across providers, so it's worth confirming whether a given provider redacts by default or only on request. What is PII redaction
Zero data retention: a configuration where a provider doesn't store submitted audio or transcripts after processing completes. It's distinct from a provider's model-training policy, since a system can retain no data for storage while still (or not) using submissions to train future models, so the two are worth checking separately.
SOC 2 Type II: an independent audit confirming a provider's security controls (access, encryption, monitoring) operated effectively over a sustained period, not just that they were designed correctly at a single point in time. It's one of the baseline checks enterprise security teams run before approving a vendor.
