AssemblyAI
AssemblyAI is an independent speech-AI company that sells speech recognition and speech-understanding APIs to developers. Dylan Fox founded it in 2017 and took it through Y Combinator's Summer 2017 batch. Its last announced funding was a $50 million Series C led by Accel in December 2023, which brought its total to $115 million. At the time it said it was serving 25 million inference calls and signing up more than 10,000 new organizations a month, with customers including Zoom, Fireflies.ai and CallRail. Where Deepgram, its closest rival, has optimized for live conversation, AssemblyAI has optimized for accuracy on hard recordings and the analysis layered on top.
Models and Products (October 2026)
| Product | What it does | List price (October 20, 2026) |
|---|---|---|
| Universal-3.5 Pro (July 7, 2026) | Batch transcription with native code-switching across 18 languages and speaker diarization built into the same model | $0.21 an hour, plus $0.02 an hour for diarization |
| Universal-3.6 Pro Realtime (September 29, 2026) | Streaming recognition with automatic detection of 32 languages | $0.45 an hour |
| Universal-Streaming | The older, cheaper streaming model | $0.15 an hour |
| Voice agent API | A complete voice-agent pipeline | $4.50 an hour ($0.075 a minute) |
| LLM Gateway (relaunched April 13, 2026) | Access to 25+ large language models for summaries, extraction and Q&A over transcripts; it replaced the earlier LeMUR framework | Per model |
Around the core models sit add-ons: keyterm prompting ($0.05 an hour), a medical mode, and PII redaction across 48 policy types that can be applied to both the transcript and the audio, which can be bleeped or silenced. A "Speaker Identification" add-on assigns names, but it infers them from what's said in the conversation rather than from voiceprints (see speaker diarization vs speaker identification).
Accuracy Claims
AssemblyAI's launch post for Universal-3.5 Pro reports a code-switching word error rate of 7.69%, against 8.77% for ElevenLabs Scribe v2 and 12.22% for Deepgram Nova-3 Multilingual. For its realtime model it reports a 537-millisecond median endpoint latency. These are vendor benchmarks, chosen and run by the vendor. Independent boards such as Hugging Face's Open ASR Leaderboard help, but their averages mix rows scored on different datasets and none matches a given product's audio, so buyers should test on their own recordings.
Deployment and How It Compares
AssemblyAI is primarily a cloud API. Self-hosted streaming exists, but its documentation puts the entry point at a $50,000 upfront commitment, whereas Deepgram offers self-hosting more broadly. AssemblyAI is cheaper than Deepgram for batch work and pricier for its newest streaming model. Its joint diarization, native code-switching and redaction suit media, meeting and research transcription, while Deepgram's turn detection and own text-to-speech suit voice agents. ElevenLabs' Scribe v2 competes on multi-speaker transcription. For teams that would rather run models themselves, open options such as Whisper, NVIDIA's Parakeet and pyannote for diarization cover the same ground with more engineering. See Deepgram vs AssemblyAI for the full comparison.
Further Reading
- Universal-3.5 Pro — AssemblyAI; the July 2026 model and its vendor benchmarks
- Universal-3.6 Pro Realtime — AssemblyAI; the September 2026 streaming model
- AssemblyAI pricing — per-hour rates and add-ons
- AssemblyAI's $50M Series C — December 2023; investors and usage figures