Last tested · By Ameen Ahmed
Best speaker diarization API: measured on public datasets
Of the five speech-to-text APIs we measured, Speech Revolutions has the lowest diarization error rate on every dataset: 16.7% pooled across three meeting datasets, against 24.6% for the next best, AssemblyAI. On VoxConverse interviews and talk shows it is 3.6% against 9.8% for AssemblyAI and 10.6% for Deepgram.
Diarization error rate (DER) is the share of audio time where the speaker label is wrong: speech that was missed, labelled where there was none, or given to the wrong person. Lower is better. Below are the full results, how they were measured, and where the gap is small.
Meeting audio: AMI, NotSoFar and DiPCo
Overlap-aware DER at a 0.25 s collar, the convention most published diarization research uses. Measured 2026-07-12 with each provider's diarization switched on. The last column is the strict figure, with no collar.
| Provider | AMI-SDM | NotSoFar | DiPCo | Pooled | Pooled, strict |
|---|---|---|---|---|---|
| Speech Revolutions Zephyr | 9.7% | 10.9% | 26.7% | 16.7% | 24.4% |
| AssemblyAI Universal 3.5 Pro | 25.3% | 23.3% | 28.9% | 24.6% | 35.9% |
| ElevenLabs Scribe V2 | 33.7% | 24.4% | 31.6% | 29.7% | 41.8% |
| Deepgram Nova-3 | 30.9% | 28.4% | 64.2% | 35.0% | 48.4% |
| Soniox V5 | 36.6% | 32.2% | 29.6% | 36.2% | 48.3% |
The lead is widest on AMI and NotSoFar, the two sets closest to everyday meeting recordings.
Interviews and talk shows: VoxConverse
Twenty recordings from the VoxConverse dev set, with two to four speakers and three to twenty minutes each. This is closer to podcast and interview audio than the meeting sets. Measured 2026-10-07.
| Provider | DER, strict | DER, 0.25 s collar |
|---|---|---|
| Speech Revolutions | 3.6% | 2.0% |
| AssemblyAI | 9.8% | 7.4% |
| Deepgram | 10.6% | 8.3% |
Most of the difference is in turn boundaries: Speech Revolutions misses two to five times less speech, and marks two to five times less silence or noise as someone talking.
How it was measured
- AMI-SDM. Recorded business meetings, four participants, one distant microphone.
- NotSoFar. Real Microsoft Teams-style meetings recorded far-field in many different rooms.
- DiPCo. Dinner-party conversations, four speakers, far-field, lots of overlap. The hardest set.
- VoxConverse. Talk shows, debates and interviews from YouTube: 2-4 speakers, 3-20 minutes each.
Each file was sent to every provider through its public API with speaker labels on, then scored against the human-labelled reference with the same scorer. Overlapping speech is scored too: where two people talk at once, missing either of them counts as an error.
The collar is a tolerance window around each speaker change. A DER quoted without its collar can't be compared with anything, which is why both are shown. Why two providers can both claim the lowest DER explains it.
Not measured: Speechmatics, Rev AI, Gladia, Google, AWS, Azure, OpenAI's gpt-4o-transcribe-diarize, and open-source pyannote or WhisperX. If you need one of them compared, the harness takes a provider adapter.
Choosing a diarization API
- Recorded meetings, interviews, podcasts, calls: speaker labels are where these products break first. A quote credited to the wrong person costs more than a misheard word. This is where the gap above matters most.
- Live audio: you need a streaming API. Speech Revolutions is batch only.
- Word accuracy: see the full benchmarks for word error rate, timestamps and multilingual results.
- Cost: diarization is included in our $0.18 an hour. Elsewhere, compare the price of an hour with it switched on.
Head to head: vs AssemblyAI, vs Deepgram. Looking to switch: AssemblyAI alternatives, Deepgram alternatives.
Questions
- What is a good diarization error rate?
- It depends on the audio and on the collar the figure was scored with. On clean recordings of two to four speakers, such as interviews, the best systems we measured are under 5%. On far-field meetings with overlapping speech, 10-25% at a 0.25 s collar is typical even for strong systems. Compare figures only when the dataset and collar match.
- Does OpenAI Whisper do speaker diarization?
- Whisper itself transcribes words but does not label speakers. Open-source pipelines such as WhisperX add diarization by pairing Whisper with pyannote. OpenAI's API has a separate gpt-4o-transcribe-diarize model that returns speaker segments. We have not benchmarked either.
- Is diarization an extra charge?
- Not with Speech Revolutions: diarization is included in the $0.003 per minute rate, as are word timestamps. Elsewhere, check whether it is priced separately, and compare the cost of an audio hour with diarization switched on.
- Can I reproduce these numbers?
- The AMI-SDM, NotSoFar and DiPCo figures come from a public harness on public datasets, so yes: run it with your own API keys.
- Which API is best for real-time diarization?
- Speech Revolutions is batch only, so it is not an option for live audio. Everything on this page measures recorded files.
Try it on your own audio
Benchmarks narrow the field; your recordings decide it. New accounts get $10 of credit, about 55 hours of audio, with no card required.