LoginSign Up
← Back to all sparks
A

AssemblyAI

COMMS
Velocity0.0

AssemblyAI stopped selling transcription and started selling the whole voice agent.

voice-agentsspeech-to-textdiarizationllm-gatewaypii-redaction
◆Current state
AssemblyAI shipped a Voice Agent API that delivers speech understanding, LLM reasoning, and voice generation over a single WebSocket at one $4.50/hr rate, built on its own models. Around it, streaming speaker diarization got an accuracy upgrade with per-word labels across US and EU regions, the LLM Gateway gained a json-repair post-processing step, and PII redaction can now return redacted and unredacted transcripts in one request.
◆Where it's heading
The supporting releases are all steps up the stack from raw transcription: an LLM gateway with output repair, speaker labels precise enough for agent turn-taking, and redaction that no longer forces a second call. The Voice Agent API is where they converge — AssemblyAI is pricing an outcome per hour rather than an API per minute, and owning the models underneath is what makes that pricing defensible.
◆Prediction
Expect the agent pipeline to gain configuration depth — interruption handling, turn detection, or telephony ingress — since a single all-in rate only holds up if the pipeline handles real call conditions. The feed also republishes launches a day apart, so entry counts here run ahead of actual releases.

◆Recent moves

  1. 4mo ago

    LLM Gateway: JSON Repair Post-Processing for Structured Output

    An optional json-repair pass fixes malformed model output before it reaches the application, enabled with one parameter. Small in scope and squarely aimed at the structured-output failure that breaks agent pipelines in production.

  2. 4mo ago

    Streaming Speaker Diarization: Major Accuracy Upgrade with Per-Word Labels

    A retrained streaming diarization model with per-word speaker labels, live in US and EU regions with no integration changes required. Word-level attribution is what a voice agent needs to know when a speaker actually changed, so this is infrastructure for the agent push rather than a standalone transcription feature.

  3. 5mo ago

    Introducing the Voice Agent API

    ⚡ SPARK

    The release the rest of this window supports: a full voice agent pipeline over one WebSocket, on AssemblyAI's own models, at one all-in hourly rate. It moves the company from a component vendor to a platform selling the finished capability.

  4. 5mo ago

    Voice Agent API (republish)

    A republished record of the Voice Agent API launch covered elsewhere in this window.

    View source ↗
  5. 5mo ago

    PII Redaction: Return Unredacted Transcripts in the Same Request

    A redact_pii_return_unredacted flag returns both versions of a transcript in one request. Minor on its face, but it removes a duplicate transcription call for anyone who needs the original text for internal review and the redacted copy for storage.

  6. 5mo ago

    PII Redaction unredacted output (republish)

    A duplicate record of the PII redaction change already covered in this window.

    View source ↗