Compare speech-to-text models at a glance
Choosing a transcription API means balancing accuracy, language coverage, real-time vs batch support, speaker diarization, and price per minute. This table puts the major speech-to-text models side by side — OpenAI Whisper, Deepgram, AssemblyAI, Rev.ai, Google, and Azure — so you can pick the right model for captions, meeting notes, call analytics, or media transcription.
How to read the table
- WER is approximate word error rate on clean English audio; lower is better.
- Languages is the number of supported languages.
- Real-time flags whether the API supports low-latency streaming for live captioning, versus batch-only async transcription.
- Diarization flags built-in speaker labelling.
- $/min is a list-price estimate per audio minute; add-on models can cost more.
Filter by real-time requirement and budget, search by model name, and click a column header to sort.
The real-time vs batch decision
This distinction shapes everything downstream. Real-time streaming transcription processes audio as it arrives, producing partial transcripts within hundreds of milliseconds. It is essential for live captions, voice agent conversations, and call centre monitoring. Batch transcription processes a completed audio file, trading latency for throughput and often lower cost per minute.
The mistake teams make is choosing a streaming provider for a batch use case (paying a latency premium they do not need) or a batch provider for a live-caption requirement (discovering the architecture will not support real-time output). The table flags the distinction clearly so you choose the right architecture first.
Diarization: when you need it and when you do not
Speaker diarization — labelling segments by speaker (“Speaker A”, “Speaker B”) — is essential for meetings, interviews, calls, and any audio with more than one participant where you need to attribute who said what. Without diarization, you get a single stream of text with no speaker labels.
It adds cost and latency, so it is worth skipping for single-speaker use cases like dictation, voiceover transcription, or lecture capture. The table flags which providers support it natively (saving you the engineering cost of building or integrating a separate diarization step).
Accuracy vs audio quality
Word error rate figures in any comparison — including this one — are measured on clean, clear audio, typically English. In practice, accuracy degrades with:
- Background noise — varies by provider but all degrade; Deepgram’s acoustic models and AssemblyAI tend to be more robust than raw Whisper on noisy input
- Accents and dialects — models trained on narrower corpora perform worse on regional accents; test on your actual audio, not generic benchmarks
- Domain vocabulary — medical, legal, and technical terms are often transcribed as similar-sounding common words without a custom vocabulary
- Audio codec and quality — phone-quality 8kHz audio produces higher WER than 16kHz or 44kHz
Always evaluate on a sample of your actual audio before committing to a provider. Benchmark numbers rarely translate directly to the audio conditions of your use case.
Tips for picking a model
- For live captions and voice agents, prioritise the real-time streaming providers — Deepgram and AssemblyAI lead on latency and accuracy.
- For bulk media or podcast transcription, batch APIs like Whisper and Rev.ai are cost-effective and accuracy is excellent on clean audio.
- For meetings and call analytics, make sure diarization and word-level timestamps are supported.
- If privacy or cost at scale dominates, self-hosting open Whisper removes per-minute fees but you pay for GPU compute and operate the pipeline yourself.
- Test on your real audio. Benchmark WERs are measured on clean reference data; your call recordings, interviews, or field audio will produce different results.