Speech-to-Text Models Compared 2026: Whisper, Deepgram, AssemblyAI, ElevenLabs Scribe

Why This Is the Opposite of Our Voice Generation Guide

Voice AI has two halves that people constantly conflate. One half turns text into audio — that is text-to-speech, and we covered it in our AI voice model showdown, which compares ElevenLabs, OpenAI, Cartesia, Hume and Sesame on how convincingly they speak.

This guide is the other half: audio in, text out. Automatic speech recognition. Transcription. The unglamorous plumbing underneath every meeting notetaker, every call-centre analytics dashboard, every voice agent that needs to know what you just said before it can answer.

The two markets barely overlap in how you evaluate them. Nobody cares about the emotional prosody of a transcript. What matters instead is word error rate on your accents, how fast the first token arrives when a caller is mid-sentence, whether the model can tell two speakers apart, and what it costs when you have 40,000 hours of call recordings sitting in cold storage.

Confusingly, several vendors now sell both. ElevenLabs built its reputation on synthesis and then shipped Scribe for recognition. OpenAI ships Whisper and TTS. Being excellent at one direction predicts almost nothing about the other, so evaluate them separately.

How We Compared

This is an editorial synthesis, not a lab test. We did not record our own audio corpus or run our own word error rate harness. What we did instead:

• Read each vendor's current official pricing page and pulled the listed per-minute or per-hour rates as of mid-July 2026.
• Read the official model and feature documentation to establish what is actually supported — streaming, diarization, timestamp granularity, language counts, file limits.
• Collected the accuracy claims each vendor publishes and attributed them to the vendor that published them, because in this market essentially every benchmark is run by a company that wins it.
• Used the community-run Open ASR Leaderboard on Hugging Face as the closest thing to a neutral reference point for open-weight models.

Where a number could not be traced to a primary source, we used a qualitative rating instead of inventing a figure. You should do the same when you evaluate: the single most useful hour you can spend is running your own audio through two free tiers.

The Landscape at a Glance

Prices are US dollars, taken from each vendor's public pricing page in July 2026. Plan tier, promotional rates and add-ons all move these numbers — verify before you budget.

1. OpenAI Whisper — Best for open-source and self-hosting

Best for: teams that cannot send audio to a third party, or that transcribe enough volume that per-minute API pricing stops making sense.

Whisper remains the reference open ASR model, and its continued relevance is mostly about licensing and portability rather than raw accuracy. The repository ships model sizes from tiny (39M parameters, roughly 1 GB VRAM) up to large (1.55B parameters, roughly 10 GB VRAM), plus a turbo variant at 809M parameters that trades translation capability for roughly 8x the speed of large. Code and weights are released under the MIT licence in the repo, and the large-v3 model card is published under Apache-2.0.

• Languages: 99, per the large-v3 model card.
• Training scale: large-v3 was trained on 1 million hours of weakly labelled audio plus 4 million hours of pseudo-labelled audio, and OpenAI reports a 10 to 20 percent reduction in errors versus large-v2.
• Timestamps: sentence-level and word-level, via the return_timestamps parameter.
• Hardware: a single consumer GPU is enough for anything up to large.
• Cost: zero licence cost — you pay in GPU hours. That flips the economics once you are transcribing thousands of hours a month.

OpenAI also hosts transcription models you can call without running anything, and the speech-to-text guide is refreshingly explicit about which model does what. GPT-4o Transcribe and GPT-4o Mini Transcribe support streaming but not word-level timestamps. Legacy whisper-1 supports word and segment timestamps plus SRT and VTT output but not streaming. A separate gpt-4o-transcribe-diarize model handles speaker separation. Per the pricing page, GPT-4o Transcribe is 0.006 US dollars per minute and the mini variant 0.003 — roughly 0.36 and 0.18 per audio hour.

Limitations: Whisper has a 30-second receptive field, so long-form audio needs chunking or a sliding window, which is where most self-host implementations lose accuracy. There is no built-in speaker diarization at all — you must bolt on a separate pipeline. And the hosted API caps file uploads at 25 MB, which is small enough to force chunking for most meeting recordings. If you are weighing self-hosting more broadly, our local LLM guide covers the same hardware trade-offs on the text side.

2. Deepgram — Best for real-time streaming and voice agents

Best for: live captioning, call centres, and voice agents where every 100 milliseconds of latency shows up as an awkward pause.

Deepgram has spent years optimising for the streaming case, and it shows in both the product surface and the price list. The current flagship is Nova-3, and the company has since added Flux, described in its docs as the first conversational speech recognition model built specifically for voice agents, with model-integrated end-of-turn detection — the model itself decides when the human has stopped talking, rather than making you tune a silence threshold.

• Pricing: the pricing page lists Nova-3 monolingual at about 0.0077 US dollars per minute for pre-recorded audio (roughly 0.46 per audio hour) and about 0.0048 per minute for streaming (roughly 0.29 per hour), with a cheaper Growth tier. Multilingual costs slightly more. Worth being blunt about this: for batch work Deepgram is the most expensive engine in this comparison. Its case rests on real-time latency, not unit price.
• Diarization: a paid add-on. Deepgram's pricing page lists speaker diarization separately at about 0.0020 US dollars per minute (roughly 0.12 per audio hour) on pay-as-you-go, so budget for it if you need speaker labels.
• Accuracy (vendor-reported): in its Nova-3 launch post, Deepgram reports a 6.84 percent median WER streaming and 5.26 percent median WER batch, claiming large margins over the next-best competitor. This is Deepgram's own benchmark on Deepgram's own test set.
• Multilingual code-switching: real-time transcription across ten languages including English, Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch, with switching mid-utterance.
• Keyterm prompting: up to 100 domain-specific terms injected at request time, no retraining.
• Free tier: 200 US dollars in credit, which buys a lot of batch hours.

Limitations: self-hosting exists but is gated. The self-hosted docs require an Enterprise plan and a product contract tied to your console project before you can pull the container images — this is not a download-and-go option. Deepgram's accuracy claims also sit awkwardly next to AssemblyAI's, which rank Nova-3 last of four on general pre-recorded speech. Both benchmarks are vendor-run.

3. AssemblyAI — Best for meetings, diarization and audio intelligence

Best for: anyone who needs more than a transcript — speaker labels, entities, topics, sentiment, summaries.

AssemblyAI has positioned itself as the speech understanding company rather than just a transcription API, and the pricing page reflects that: a base rate plus an a-la-carte menu of intelligence features.

• Pricing: Universal-3.5 Pro is 0.21 US dollars per hour for pre-recorded audio; the older Universal-2 is 0.15 per hour. Streaming is 0.45 per hour base for Universal-3.5 Pro Realtime, or 0.15 per hour for Universal-Streaming in English or multilingual.
• Diarization: an add-on, but a cheap one — 0.02 per hour on top of async transcription, 0.12 per hour for streaming.
• Audio intelligence add-ons: entity detection 0.08 per hour, topic detection 0.15, translation 0.06, custom formatting 0.03, sentiment analysis 0.02. These stack additively on the base rate.
• Accuracy (vendor-reported): on AssemblyAI's own benchmarks page, Universal-3.5 Pro scores 4.35 percent WER on general pre-recorded speech, ahead of OpenAI GPT-4o Transcribe (5.338 percent), ElevenLabs Scribe v2 (5.869 percent) and Deepgram Nova-3 (6.662 percent). The streaming model posts 5.53 percent, with a median time-to-complete-turn of 335 milliseconds against a competitor range of 247 to 1,136 milliseconds.
• Languages: 18 for Universal-3.5 Pro, 99 for the older Universal-2 — worth knowing, because the newest model is narrower.
• Free tier: 50 US dollars in credits, no card required.

Limitations: the billing model for streaming has a sharp edge. AssemblyAI bills streaming by session duration, not audio duration — a WebSocket held open for 60 minutes while only 30 minutes of audio is sent bills for 60 minutes. If your client keeps sockets alive between utterances, your bill will be roughly double what you modelled. There is also no self-hosting, and the flagship model's 18-language coverage is much narrower than Whisper's or Scribe's.

If you want the finished product rather than the API, most of the tools in our AI meeting notetaker comparison are wrappers around engines like this one.

4. ElevenLabs Scribe — Best for multilingual and media workflows

Best for: subtitling, dubbing prep, podcast production, and any pipeline where the audio is not in English.

Scribe is ElevenLabs turning its voice expertise around. The pitch is breadth of language coverage plus media-friendly output — the kinds of details that matter when you are building subtitles rather than a voice agent.

• Languages: 90-plus, tiered by expected accuracy. ElevenLabs publicly buckets them: 34 languages at 5 percent WER or better, 20 in the 5 to 10 percent band, 18 in the 10 to 20 percent band, and 18 more from 20 to 50 percent. Publishing the weak tiers is more honest than most vendors bother to be.
• Diarization: up to 48 distinct speakers in batch mode, with per-segment timestamps — the highest speaker ceiling of anything here.
• Audio event tagging: detects non-speech events such as laughter and footsteps, which is genuinely useful for subtitle tracks.
• Realtime: Scribe v2 Realtime is marketed at roughly 150 milliseconds latency across 90-plus languages, with automatic language detection and mid-conversation language switching.
• Pricing: the API pricing page lists Scribe at 0.22 US dollars per hour of audio and Scribe v2 Realtime at 0.39 per hour, with entity detection at 0.070 per hour and keyterm prompting at 0.050 per hour. Note the product page quotes from 0.40 per hour, so the effective rate depends on your plan tier — check your own account before budgeting.

Limitations: the realtime variant does not currently do speaker diarization — if you need live speaker labels, use Deepgram or AssemblyAI instead and keep Scribe for batch. The two ElevenLabs pages we checked also quote different starting prices, which is a mild red flag for cost modelling. And ElevenLabs is fundamentally a voice-generation company; ASR is the newer, smaller side of the house.

5. NVIDIA Parakeet TDT 0.6B v3 — Best for cheap self-hosted throughput

Best for: bulk transcription of a large archive on your own hardware, where speed per dollar dominates.

Parakeet is the model to reach for when Whisper is too slow. It is a 600-million-parameter model released under CC-BY-4.0, covering 25 European languages, and its headline number is throughput rather than accuracy.

• Speed: the model card reports an RTFx of 3,332.74 on the Open ASR Leaderboard — meaning it transcribes far, far faster than real time on suitable hardware.
• Accuracy: an average WER of 6.32 percent across the leaderboard's datasets, with 1.93 percent on LibriSpeech test-clean and 3.59 percent on test-other.
• Timestamps: word, segment and character level.
• Footprint: at least 2 GB of RAM to load the model; runs via NeMo or Hugging Face Transformers on Volta, Ampere, Hopper or Blackwell GPUs.

Limitations: 25 European languages is a much narrower net than Whisper's 99, there is no built-in diarization, and you are responsible for the whole serving stack. The same hardware maths that governs running LLMs locally applies here.

What the Benchmarks Actually Say

Here is the uncomfortable truth about ASR accuracy in 2026: the leaderboards are vendor-run and they contradict each other.

AssemblyAI publishes a benchmark where AssemblyAI wins and Deepgram comes last. Deepgram publishes a benchmark where Deepgram wins by a huge margin. ElevenLabs says Scribe achieves the lowest word error rate recorded on industry-standard benchmarks. These claims cannot all be true on the same audio, and they are not lying — they are each choosing test sets that flatter their own training distribution.

Treat that table as one vendor's data point, not as ground truth. The only benchmark that matters is yours. Take an hour of representative audio — the accents you actually serve, the jargon you actually use, the phone-quality compression you actually receive — and push it through the free tiers. Deepgram gives 200 US dollars in credit; AssemblyAI gives 50. That is more than enough to settle the argument empirically.

This is the same evaluation discipline we apply to multimodal models and to inference providers: published numbers set your shortlist, your own data picks the winner.

Which Should You Choose?

Best raw accuracy on pre-recorded audio: AssemblyAI Universal-3.5 Pro, with ElevenLabs Scribe v2 close behind. Both are in the sub-6-percent WER band on general speech according to the benchmarks each publishes. At 0.21 US dollars per audio hour, AssemblyAI is not meaningfully more expensive than the alternatives.

Best real-time and streaming: Deepgram. Nova-3 for general live transcription, Flux when you are building a voice agent and want end-of-turn detection baked into the model rather than hacked on with a silence timer. AssemblyAI's Universal-3.5 Pro Realtime is the credible alternative if you also want its intelligence features — just watch the session-duration billing.

Best open-source and self-hostable: OpenAI Whisper for breadth (99 languages, permissive licence, runs anywhere), NVIDIA Parakeet when throughput is the constraint and your languages are European. Neither gives you diarization out of the box.

Best for meetings and diarization: AssemblyAI, because speaker labels sit alongside entity detection, topic detection and sentiment in the same request. Deepgram is a strong second and includes diarization for free. ElevenLabs Scribe wins on sheer speaker count if you are transcribing a panel or a roundtable.

Best value: AssemblyAI Universal-2 at 0.15 US dollars per hour is the cheapest listed batch API in this comparison. But past a few thousand hours a month, self-hosted Whisper or Parakeet on your own GPUs will beat every API on unit cost — at the price of running the infrastructure yourself.

Best multilingual: Whisper (99 languages) if self-hosting, ElevenLabs Scribe (90-plus, with published accuracy tiers) if not. Note that AssemblyAI's newest flagship covers only 18 languages, so the newest model is not automatically the broadest.

Conclusion

Speech-to-text has quietly become a commodity at the bottom of the market and a genuine differentiator at the top. Batch transcription of clean English audio is solved and cheap — any of these engines will do it for under half a dollar an hour. The remaining hard problems are the ones the marketing pages gloss over: low-latency streaming that does not clip the start of a sentence, diarization that survives crosstalk, and accuracy on accented, noisy, jargon-heavy audio that looks nothing like LibriSpeech.

Pick Deepgram if latency is the constraint, AssemblyAI if understanding is, Whisper or Parakeet if your audio cannot leave the building, and Scribe if the audio is not in English. Then ignore all of that and run your own hour of audio through two free tiers, because the vendor benchmarks disagree with each other and only your data can break the tie.

And if you came here looking for the other direction — making machines talk rather than listen — that is our voice generation comparison.

This is an editorial synthesis of vendor documentation and public sources; see our [methodology](/methodology). Verify current details with each provider.