Best Speech-to-Text API 2026: Scribe v2 vs StepAudio vs AssemblyAI vs Deepgram

On Artificial Analysis's independent AA-WER v2 leaderboard, captured 16 October 2026, StepFun's StepAudio 3 ASR scores 1.7% word error rate, ElevenLabs Scribe v2 2.2%, AssemblyAI Universal-3 Pro 3.1% and Deepgram Nova-3 5.2%. All four are among 60 models in the table. Every vendor also publishes its own comparison, and those disagree with the independent ranking.

This post separates the independent numbers from the company-reported ones, shows the list-price math, and flags where the evidence is thin. Streaming latency is not compared here because the sources use different methods.

Which speech-to-text API is most accurate on independent tests?

By Artificial Analysis's AA-WER v2, StepAudio 3 ASR is first among the four, with Scribe v2 second, Universal-3 Pro third and Nova-3 fourth. Counting down the table, Scribe v2 sits 4th of 60 and Nova-3 50th of 60 by rounded score; Fun-Realtime-ASR-preview ties StepAudio 3 ASR at 1.7%.

All rows below are independent measurements from Artificial Analysis, benchmark AA-WER v2 (normalizer version v2.2), non-streaming, captured 16 October 2026. Prices are Artificial Analysis's normalised estimate per 1,000 minutes.

One version caveat: the leaderboard row is labelled Universal-3 Pro. AssemblyAI says its async flagship is now Universal-3.5 Pro. The sources do not say whether the leaderboard row has been re-run on 3.5 Pro, so treat the 3.1% as belonging to the model name shown.

What does AA-WER v2 measure, and how is it scored?

AA-WER v2 is a word error rate, (substitutions + insertions + deletions) divided by reference words, averaged by audio duration across about 8 hours from three datasets. The weights are 50% AA-AgentTalk, 25% VoxPopuli-Cleaned-AA and 25% Earnings22-Cleaned-AA.

The dataset sizes are small for the long-form component:

Both reference and hypothesis text pass through OpenAI's Whisper normalizer, plus Artificial Analysis's own additions for digit formatting, spelled-out letters, URLs and spelling variants. Where a model accepts a prompt, the prompt is "Transcribe the audio verbatim, outputting only spoken words in sequence." Artificial Analysis says it tests API endpoints directly.

The methodology has moved. AA-WER v2.0 arrived in February 2026, replacing a v1.0 that used AMI SDM, VoxPopuli and Earnings22. Scores from the two versions should not be treated as comparable (an inference, since the methodology changed datasets, weights and normalizer), and v2.1 (April 2026) and v2.2 (May 2026) updated the normalizer again.

Who runs AA-WER, and does it have conflicts or contamination risk?

Artificial Analysis operates the benchmark and the leaderboard, and says it runs independent WER testing across all model-provider pairs. It is a third-party evaluator, not one of the four vendors compared here.

On contamination, Artificial Analysis states that public test sets carry a growing risk of data contamination and contained errors in ground-truth transcripts. Its answer is a held-out proprietary set (AA-AgentTalk) and cleaned references for the two public sets.

The trade-off is auditability. The half of the score that carries the most weight cannot be inspected or rerun by outsiders. The leaderboard publishes point estimates only, so no one can tell from it whether 2.2% versus 3.1% is outside sampling noise. That is an inference from what is published, not a statement from Artificial Analysis.

What do the vendors claim about their own models?

Each vendor reports results on its own suites. All of these are company claims.

Two details limit these claims. ElevenLabs's accuracy chart compares Scribe v2 only with GPT 4o Transcribe, Gemini 2.5 Pro and Deepgram Nova 3, with no numeric values on the page. StepFun's page gives relative error reductions without absolute English WER or dataset names.

Deepgram says it tested Nova-3 on 2,703 audio files totalling 81.69 hours and reports a median streaming WER of 6.84%. The page is undated, and the comparison models are older ones.

Where do vendor and independent results disagree?

The clearest case is ElevenLabs versus AssemblyAI. AssemblyAI reports its flagship ahead of Scribe v2 on code-switching WER (7.69 vs 8.77), voice-agent WER (5.19% vs 7.78%), entity error rate (14.4% vs 18.5%) and diarization cpWER (30.17 vs 35.26). The independent AA-WER v2 has Scribe v2 at 2.2% and Universal-3 Pro at 3.1%.

These are not necessarily contradictory, and the gap is consistent with different test sets. AssemblyAI's voice-agent WER and entity-error figures are for its Universal-3.6 Pro Realtime model, not the batch Universal-3.5 Pro. The post's author, who leads Voice AI at AssemblyAI, writes that it is not a neutral referee's writeup. The same post says Scribe v2 leads on some public benchmarks.

For Deepgram, AssemblyAI's suite puts Nova-3 at 8.64% on voice-agent audio, and AA-WER v2 puts it at 5.2% overall. Different datasets and different metrics, so the numbers cannot be compared directly. What both agree on is the ordering: Nova-3 behind the ElevenLabs and AssemblyAI models.

Normalisation is a quieter source of disagreement. An arXiv study (2609.21084) on Polish speech says in its abstract that lack of number normalization can shift WER by more than 2 percentage points, often more than the gaps between systems in popular multilingual benchmarks. It is a Polish study, not a test of Artificial Analysis's English normalizer, but it shows why two WER tables on different normalizers should not be mixed. The same logic applies to harness choices in other eval domains and to saturation as scores approach the floor.

What does each API cost, and what does it cost per 1,000 hours?

List prices cluster between $0.21 and $0.26 per hour for batch transcription as of 16 October 2026. The per-1,000-hour figures below are straight multiplication of the published hourly rate.

Add-ons change the total. AssemblyAI lists keyterms prompting at +$0.05/hr and standard diarization at +$0.02/hr. Its Realtime model is $0.45/hr base. Cartesia, which competes with ElevenLabs in voice AI, reports $0.39/hr for Scribe v2 Realtime. Deepgram lists $0.0077/min for Nova-3 streaming pay-as-you-go.

On free tiers, AssemblyAI offers $50 in credits, good for up to 185 hours of pre-recorded audio, and Deepgram's guide lists $200 in credits. The Deepgram rates come from a guide with a 2025 slug, so confirm them on Deepgram's current pricing page before budgeting.

Note the StepFun gap. The docs say $0.24 per hour, which is $4.00 per 1,000 minutes, while Artificial Analysis shows $6.67, or $0.40 per hour. Artificial Analysis says it converts processing-time, token or subscription billing into a common unit, but neither source explains this particular difference.

How fast are they, and what do the language limits look like?

Deepgram Nova-3 is the fastest on Artificial Analysis's batch measure. Its speed factor is 601.2 audio seconds per second, against 108.6 for AssemblyAI Universal-3 Pro, 73.7 for StepAudio 3 ASR and 65.6 for Scribe v2. Artificial Analysis's speed factor is a median over the last 7 days on 10-minute audio and includes network latency.

Speed here is batch throughput, not time to first word. For streaming, vendors report different things. ElevenLabs claims under 150 ms for Scribe v2 Realtime, while AssemblyAI reports a 307 ms median time to final transcript measured with Pipecat. Those are not the same measurement and should not be ranked against each other.

Language coverage, as claimed:

StepAudio 3 ASR Max also differs in interface. It uses HTTP plus SSE rather than a bidirectional socket, and StepFun lists a separate stepaudio-2.5-asr-stream model for real-time bidirectional streaming.

Which API should you pick for which job?

Match the choice to the constraint, then run your own audio. The published evidence supports these tentative readings:

• Accuracy on English batch audio: StepAudio 3 ASR and Scribe v2 lead the independent table. ElevenLabs claims 90+ languages, while StepAudio's languages beyond Chinese and English are in preview.
• Throughput and cost per hour of audio: Nova-3 is fastest on the independent speed measure, but sits 50th of 60 on accuracy. AssemblyAI's list price is the lowest of the four.
• Voice agents and entity accuracy: AssemblyAI's entity-error numbers are the only ones published on this, and they are company-reported on a proprietary suite.
• Chinese-English mixed speech: StepFun's own benchmarks emphasise this, but AA-WER v2 is English and does not test it.

A practical step: sample 2 to 5 hours of your own production audio, transcribe it with each API using the same normalizer, and compare entity and keyword errors as well as WER.

What's still unclear?

• Which AssemblyAI model is on the leaderboard. The leaderboard row says Universal-3 Pro, while the vendor's current async flagship is Universal-3.5 Pro.
• StepFun's price. The docs and Artificial Analysis differ by $0.16 per hour, and neither source explains why.
• Confidence intervals. The independent leaderboard gives point estimates. With 469 short agent clips, 628 parliamentary clips and 6 long calls, gaps under a few tenths of a point are probably not decisive. That is an inference.
• Ties. StepAudio 3 ASR and Fun-Realtime-ASR-preview both show 1.7% after rounding.
• English-only scope. The score weights English speech. It does not support claims about the 90+ language counts or the preview languages.
• Vendor suites. AssemblyAI's, Deepgram's, ElevenLabs's and StepFun's comparisons use proprietary or undisclosed test sets, so no one can reproduce them from the published pages.
• Docs quality, funding and per-task latency percentiles are not reported in the public sources cited here, so this post does not rate them.

Last checked: 16 October 2026.