Deepgram vs AssemblyAI

Accuracy and the Benchmark Problem

Both companies claim to be the most accurate, and both publish the numbers to prove it. 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, 12.22% for Deepgram Nova-3 Multilingual and much worse for OpenAI's GPT-4o Transcribe. Deepgram's own launch claims center on speed, such as Flux's sub-400 ms turn decisions. These are vendor numbers, chosen and run by the vendor, and should be read that way.

The independent signal is thinner than it looks. Hugging Face's Open ASR Leaderboard is the best-known public board, but its averages mix rows scored on different sets of datasets, and no public test matches a given product's audio. Accuracy depends on that audio: accents, crosstalk, phone codecs, domain vocabulary. Both vendors offer keyterm prompting to bias recognition toward names and jargon (Deepgram up to about 100 terms per request on Nova-3 and Flux; AssemblyAI as a $0.05/hr add-on), and that often matters more than a point of benchmark WER. The only reliable test is a few hours of your own audio, scored blind. For open models, see Whisper vs Parakeet.

Real-Time Speech and Voice Agents

Deepgram's clearest lead is in conversational AI. Flux is a streaming recognizer that also decides when a speaker has finished, using a model rather than a silence timer, which lets an agent reply quickly without cutting people off. Deepgram says Flux makes that call in under 400 ms. Flux Multilingual added Spanish, French, German, Hindi, Russian, Portuguese, Japanese, Italian and Dutch in April 2026, with switching between languages mid-call. Deepgram also has its own text-to-speech, Flux TTS, launched on 2026-08-12, so one vendor can cover both ends of a voice agent.

AssemblyAI has closed much of the gap. Universal-3.6 Pro Realtime detects 32 languages automatically, and AssemblyAI reports a 537 ms median endpoint latency and strong results on third-party agent benchmarks such as Pipecat's. Its streaming price, $0.45 an hour, is higher than Deepgram's Nova-3 streaming rate; the older Universal-Streaming remains at $0.15 an hour. Both companies now sell a complete voice-agent API at the same pay-as-you-go price of $0.075 a minute, so the decision for agent builders comes down to latency on their own calls, language coverage and the surrounding tooling.

Diarization and Multilingual Audio

For recordings with several speakers, AssemblyAI has the more integrated design. Universal-3.5 Pro recognizes words and speakers in one joint model rather than running a separate diarizer afterward, and it supports code-switching natively across 18 languages, which matters for interviews and calls where people switch languages mid-sentence. Diarization is still billed separately, at $0.02 an hour for pre-recorded audio. AssemblyAI also sells a "Speaker Identification" add-on, but it infers names from the conversation's content, not from voiceprints (see Speaker Diarization vs Speaker Identification).

Deepgram supports diarization on its Nova batch models and on streaming, where it lists diarization at $0.002 a minute. Its documentation does not describe diarization for Flux, which is built for two-party agent calls rather than multi-speaker recordings. For podcast, meeting and media transcription with many voices, AssemblyAI is the stronger default; for a live call between one person and an agent, the question does not arise.

Pricing in Practice

At list prices on 2026-10-20, AssemblyAI is cheaper for batch transcription: Universal-3.5 Pro costs $0.21 an hour against about $0.26 an hour for Deepgram Nova-3 on pay-as-you-go, though Deepgram's Growth tier brings Nova-3 down to about $0.22. Add-ons change the math. AssemblyAI prices diarization, keyterms, PII redaction (audio and text separately) and medical mode as per-hour extras; Deepgram prices diarization, redaction and keyterm prompting as per-minute extras on streaming. A pipeline that needs diarization and redaction should price the full bundle, not the headline rate.

For streaming the order flips: Deepgram Nova-3 streaming lists at about $0.29 an hour and Flux English at about $0.39, against $0.45 for AssemblyAI's newest realtime model. New accounts get starter credit on both, $200 at Deepgram and $50 at AssemblyAI. Deepgram's April 2026 announcement described a limited-time promotional rate for streaming, so check current pricing before committing volume.

Deployment, Privacy and Compliance

Deepgram offers more ways to keep audio in-house. Its models can run self-hosted in Docker or Kubernetes on an Enterprise plan, or air-gapped through Amazon SageMaker, and it runs regional endpoints in the EU, Australia and, since 2026-09-15, India. Its October 2026 self-hosted release dropped Whisper and made watermarking of Flux TTS output mandatory. That flexibility suits contact centers, healthcare and public-sector buyers with data-sovereignty rules.

AssemblyAI is primarily a cloud API. It offers self-hosted streaming, but its documentation puts the entry point at a $50,000 upfront commitment. Its strength on the compliance side is redaction: PII policies cover 48 entity types and can be applied to both the transcript and the audio itself, which can be bleeped or silenced. Deepgram's entity redaction, by contrast, covers 50+ types but in English only, with number redaction across all languages.

Beyond the Transcript

The two have diverged on what comes after transcription. AssemblyAI retired its LeMUR framework in favor of an LLM Gateway, relaunched on 2026-04-13 with more than 25 large language models, so summaries, extraction and question answering over transcripts can run through the same account. Deepgram has stayed focused on speech in and speech out, adding text-to-speech and the voice-agent stack instead. Teams that already have an LLM provider will not miss AssemblyAI's gateway; teams that want one bill for speech and analysis may prefer it.