Every transcription vendor quotes an accuracy figure, and the figures cluster suspiciously close together. They are measured on curated benchmark audio: one speaker, good microphone, no crosstalk, standard accent, general vocabulary. Almost nothing you need transcribed looks like that.
Real recordings have people talking over each other, product names no model has heard, background noise, and a range of accents. Performance on those is what matters and it is not what gets published.
What changed in 2026
- Open-weight models became genuinely competitive. Self-hostable transcription models closed most of the gap with commercial services, which changed the calculus for privacy-sensitive audio.
- Diarization improved unevenly. Speaker separation got better in clean conditions and remained the weak point in overlapping speech.
- Real-time latency dropped. Streaming transcription became responsive enough for live captioning and conversational applications.
- Vocabulary biasing became standard. Supplying expected terms and names as a hint moved from an advanced feature to a common one.
What actually determines quality
| Factor |
Impact |
Can you control it |
| Microphone quality and placement |
Very high |
Yes; largest single lever |
| Overlapping speech |
Very high |
Partly, through meeting etiquette |
| Domain vocabulary and names |
High |
Yes; supply a term list |
| Accent and dialect coverage |
High |
No; test with your speakers |
| Background noise |
High |
Partly |
| Streaming versus batch |
Moderate |
Yes; batch is more accurate |
| Audio codec and bitrate |
Moderate |
Often yes |
| Model choice |
Moderate |
Yes, and less than the above |
Microphone quality outranks model choice, which is not what people expect. A good model on poor audio underperforms an average model on clean audio, and improving the recording setup is usually cheaper than switching vendors.
Vocabulary biasing is the highest-value software lever. Supplying a list of expected names, product terms, and acronyms before transcription measurably improves accuracy on exactly the words that matter most in a business recording — and those are the errors people actually notice.
Testing it properly
Collect twenty recordings that represent your hard cases: the noisy conference room, the speaker with the accent your team keeps mis-transcribing, the call with three people interrupting. Not your clean samples.
Transcribe them by hand or correct machine output carefully to produce ground truth. This is tedious and there is no substitute.
Then score, and score on what matters rather than on overall word error. A transcript that misses filler words is fine; one that gets a customer name or a dollar figure wrong is not. Weight errors on names, numbers, and domain terms more heavily than general words, because that reflects how the output will actually be used.
Test streaming separately from batch if you need both. A streaming model commits to words before hearing what follows, so it cannot revise based on later context, and its accuracy on the same audio is meaningfully lower.
Finally, consider where the audio goes. Voice recordings are among the more sensitive data types, and self-hosted open models are now good enough that keeping audio in-house is a genuine option — relevant if the recordings touch anything covered in health data privacy guide.
Common mistakes
- Choosing on published word error rate. Measured on audio unlike yours.
- Ignoring the microphone. Bigger lever than the model.
- Skipping vocabulary hints. Free accuracy on the words that matter most.
- Testing on clean samples. Measures the case you did not need help with.
- Assuming diarization works in crosstalk. It is the persistent weak point.
FAQ
Are open-weight models good enough now?
For many workloads, yes, and they keep audio under your control. Commercial services still lead on diarization and on turnkey features.
How much does accented speech affect accuracy?
Meaningfully, and it varies by model and accent. This is precisely why testing with your actual speakers rather than a benchmark matters.
Can I improve accuracy without changing vendors?
Usually. Better microphones, vocabulary hints, and reducing crosstalk typically produce a larger improvement than switching models.
Is real-time worse than batch?
Yes, structurally. Streaming cannot use future context to revise earlier words. Use batch wherever latency permits.
Where to go next
For the output side of voice systems, read TTS voices compared and realtime voice API. For voice as an authentication factor, voice print authentication.