AssemblyAI says Sync HTTP has a niche role beside streaming STT
AssemblyAI says most voice agents should keep using streaming speech-to-text, while its Sync HTTP API should be reserved for narrower jobs such as second-pass verification, recovery after a dropped WebSocket, or custom low-volume systems that already handle turn detection. The company said in a blog post published on 2026-10-30 that Universal-3.6 Pro Realtime remains the default starting point because it returns partial transcripts during a caller's turn and manages turn-ending logic that Sync leaves to the client.
Why streaming remains the default
According to AssemblyAI, a streaming session can return partial transcripts while a caller is still speaking, detect when a turn is complete, and hold the turn open around entities such as phone numbers and order IDs so they are not split. The company said the stream also provides a SpeechStarted signal that can be used to stop text-to-speech output when a caller begins talking. By contrast, Sync uses a single HTTP request per finished clip and returns one transcript after the clip ends, with no socket, partials, or built-in turn detection.
The post frames that difference as an implementation choice rather than a universal performance verdict. For frameworks such as LiveKit or Pipecat, AssemblyAI said the streaming path matches common voice-agent architectures, while using Sync as the primary STT would require custom orchestration around voice-activity detection and endpointing.
Where Sync fits
AssemblyAI described three cases where Sync can still be useful. First, it can act as a second pass on low-confidence turns or high-stakes values such as card numbers, account IDs, or spelled-out names. In that pattern, the company recommends keeping streaming as the primary transcript and sending the finished turn audio to Sync only when a re-check is needed. The post says Sync returns the finished transcript in about 134 ms at p50, which is fast enough to add little delay to the turn.
Second, Sync can recover audio after a streaming connection drops. AssemblyAI said a network interruption can leave several seconds of audio untranscribed, and streaming cannot be replayed faster than real time because throughput is throttled at about 1.25 times real time. In its example, the client keeps a rolling buffer of recent audio, sends the audio since the last finalized turn to Sync after a disconnect, then reconnects the stream.
Third, Sync can replace the STT stage entirely for teams that already own their own turn detection and run low-volume agents. In that model, each turn becomes one HTTP request, with the client deciding when speech starts and ends before posting the clip to Sync.
Pricing and operational trade-offs
The post says Sync is priced at $0.30 per hour of audio sent, while Universal-3.6 Pro Realtime is priced at $0.45 per hour. AssemblyAI also said Sync accepts audio clips from 80 milliseconds up to 20 minutes per request, and that SDKs and the live-upload endpoint accept up to 120 seconds of WAV or raw PCM. It added that teams using Sync as the main STT should pre-warm the connection with GET /v1/warm so the TLS handshake is not on the critical path of the first turn.
AssemblyAI's broader material says Universal-3.6 Pro Realtime is a streaming speech-to-text model for voice agents, not a complete speech-to-speech agent, and that custom applications typically provide their own LLM and text-to-speech layers. Related company material published on 2026-09-29 said the model supports 32 languages, automatic language detection and native mid-sentence code-switching. AssemblyAI's post also referenced vendor-reported benchmark figures, including a September 28, 2026 Pipecat benchmark snapshot with 0.96% pooled semantic word error rate and 307 ms median latency, alongside a reported 5.19% word error rate on its own English voice-agent benchmark. Those figures were presented as benchmark-specific and not as guarantees of production performance.
The company's conclusion is narrow: use streaming for nearly every voice-agent turn, add Sync where a turn needs a second look or a connection needs rescuing, and run the full agent on Sync only when the team already owns turn detection and the volume is low enough for the savings to matter.