Best speech-to-text (STT) models in 2026
Short answer: Best depends on the metric. For ordinary transcript accuracy (most accurate by WER), universal-3.5-pro from AssemblyAI currently leads the rolling 7d benchmark at 3.3% WER. For voice agent workflows that must recover exact values (best Task Success Rate), Deepgram Nova-3 leads at 71.0%, but the top results are close enough that the workflow-specific entity rows matter more than a universal winner.
“Best” is not one ranking. Most accurate transcripts, lowest finalization latency, and best exact-value recovery are measured on separate tables: a live Word Error Rate and latency benchmark, and a separate exact structured-value benchmark. The values are not blended into a made-up overall score.
Most accurate, lowest latency, or best exact-value recovery
Start with the job your transcript must do. Each row names the metric that should decide “best,” then points to the live ranking.
| Use case | Read this metric | Evidence |
|---|---|---|
| Captions, notes, search, analytics | Most accurate (lowest WER) | Read the WER ranking in [01] ↓ |
| Forms, callbacks, workflow automation | Best Task Success Rate | Read the TSR ranking in [02] ↓ |
| Individual emails, phones, paths, and IDs | Best CTEM for the specific entity type | Read the full entity matrix → |
| Realtime exact-value capture | Best streaming TSR | Compare streaming rows in [02] ↓ |
| Lowest-latency transcript finalization | Lowest latency (p95 TTFS, then median) | Compare the latency columns in [01] ↓ |
Most accurate speech-to-text (STT) models by Word Error Rate
Lower WER is better. These figures are Coval's rolling 7d aggregate on identical audio, mirrored with attribution and last synced Aug 14, 2026. Median and p95 TTFS show lowest-latency finalization, not the time to the first partial transcript.
| Rank | Model | Provider | WER | Median TTFS | p95 TTFS |
|---|---|---|---|---|---|
| 1 | universal-3.5-pro | AssemblyAI | 3.3% | 124 ms | 280 ms |
| 2 | realtime | Reson8 | 3.6% | 269 ms | 405 ms |
| 3 | chirp_3 | 4.2% | 805 ms | 1,120 ms | |
| 4 | enhanced | Speechmatics | 4.4% | 326 ms | 515 ms |
| 5 | grok-stt | xAI | 4.5% | 192 ms | 255 ms |
| 6 | gpt-4o-transcribe | OpenAI | 4.7% | 713 ms | 1,059 ms |
| 7 | inworld-stt-1 | Inworld AI | 4.8% | 98 ms | 121 ms |
| 8 | ink-2 | Cartesia | 5.0% | 105 ms | 137 ms |
| 9 | chirp_2 | 5.1% | 763 ms | 1,179 ms | |
| 10 | pulse | Smallest | 5.2% | 192 ms | 293 ms |
The current accuracy-latency Pareto frontier contains 4 models: parakeet-tdt-0.6b-v3 (Together AI), stt-rt-v5 (Soniox), inworld-stt-1 (Inworld AI), universal-3.5-pro (AssemblyAI). A frontier model is not beaten by another model on both WER and median TTFS. Open the complete live STT benchmark →
Most accurate speech-to-text (STT) models for exact values
This independent snapshot uses 300 English workplace recordings containing 1,482 audited values across 26 types. TSR requires every target value in a recording to be correct; CTEM scores individual values. Scores are fractions rendered as percentages, and higher is better.
| Rank | Model | Provider | Mode | TSR | CTEM | Format-invariant WER |
|---|---|---|---|---|---|---|
| 1 | Deepgram Nova-3 | Deepgram | batch | 71.0% | 91.4% | 8.2% |
| 2 | ElevenLabs Scribe v2 | ElevenLabs | batch | 68.0% | 91.8% | 6.8% |
| 3 | Deepgram Nova-3 | Deepgram | streaming | 61.0% | 88.3% | 9.5% |
| 4 | ElevenLabs Scribe v2 Realtime | ElevenLabs | streaming | 59.0% | 88.1% | 5.1% |
| 5 | Google Cloud Chirp 3 | Google Cloud | batch | 58.0% | 88.7% | 5.8% |
| 6 | OpenAI GPT Realtime (Whisper) | OpenAI | streaming | 57.3% | 88.7% | 8.6% |
| 7 | Google Cloud Chirp 3 | Google Cloud | streaming | 56.3% | 87.3% | 6.8% |
| 8 | Inkling (Thinking Machines) | Thinking Machines Lab | batch | 54.7% | 85.3% | 5.0% |
| 9 | OpenAI GPT-4o Transcribe | OpenAI | batch | 54.0% | 87.0% | 4.6% |
| 10 | Whisper large-v3 | OpenAI | batch | 53.3% | 87.3% | 5.9% |
Deepgram Nova-3 and ElevenLabs Scribe v2 are separated by 3.0 percentage points on TSR. With only 300 clips, treat that as an overlapping top group rather than a durable universal ordering. Check the per-entity matrix for the values your application uses. Open all systems and entity types →
What each ranking measures
The WER/latency table and structured-value table use different audio, protocols, windows, and scopes. Their percentages cannot be averaged or compared directly. This page combines the decision, not the scores.
| Benchmark | What it answers | Primary metrics | Provenance |
|---|---|---|---|
| Live STT benchmark | Which model has the fewest word errors, and how quickly does it finalize? | Average WER, median and p95 TTFS | Coval sells voice-agent evaluation infrastructure. Rolling data and open runner are attributed to Coval. |
| Structured STT benchmark | Does the transcript preserve every exact value a downstream task needs? | TSR, CTEM, format-invariant WER | Dataset by Besimple AI; baselines run and scored by Openbenchmarks. |
Reproduce the data: Coval TTS/STT benchmark mirror · structured STT audio, transcripts, decisions, and scoring code.
Best speech-to-text (STT) models: common questions
What is the best speech-to-text (STT) model in 2026?
Best depends on the metric. For ordinary transcript accuracy (most accurate by WER), universal-3.5-pro from AssemblyAI currently leads the rolling 7d benchmark at 3.3% WER. For voice agent workflows that must recover exact values (best Task Success Rate), Deepgram Nova-3 leads at 71.0%, but the top results are close enough that the workflow-specific entity rows matter more than a universal winner.
What is the most accurate speech-to-text (STT) model?
Most accurate means best WER for readable transcripts, or best Task Success Rate when a wrong digit, email, path, or ID breaks the workflow. Those are different rankings; use the tables below instead of one blended score.
Which is the best STT model provider?
No provider wins every definition of accuracy. AssemblyAI supplies the current lowest-WER model in the rolling transcript benchmark, while Deepgram supplies the highest-TSR model for exact structured values. Choose the provider whose model wins on your audio, deployment mode, language, latency budget, and failure-sensitive fields.
Is the model with the lowest Word Error Rate always the most accurate?
Only for a task defined by average word correctness. On the structured benchmark, AssemblyAI Universal 3.5 Pro has the lowest format-invariant WER at 3.9%, while Deepgram Nova-3 leads recording-level Task Success Rate at 71.0%. A transcript can have few wrong words and still corrupt the one phone number, email, path, or ID a workflow needs.
Which speech-to-text (STT) model has the lowest latency?
Lowest latency here means lowest p95 Time to Final Segment, not the first partial transcript. Tiny median gaps at the fast end are usually smaller than the rest of a voice pipeline. Use the TTFS columns in the WER table, then confirm on the live STT board.
What is the best speech-to-text (STT) model for realtime voice agents?
Deepgram Nova-3 currently has the highest structured-value Task Success Rate among streaming systems on this board at 61.0%. That is not a complete voice-agent recommendation: also compare p95 finalization latency, interruption handling, endpointing, language coverage, and performance on your own calls.
How should I compare speech-to-text (STT) models fairly?
Send identical audio to every model, separate batch from streaming, and report at least average WER, p95 finalization latency, and a task-specific metric for critical values. Keep prompts, vocabulary boosts, cleanup, and endpointing settings controlled or disclose them. Finally, repeat the test on production audio because accents, noise, languages, and domain vocabulary can change the ordering.