LoginSign Up
← Back to all sparks
D

Deepgram

COMMS
Velocity0.0

Deepgram is widening language coverage while quietly replacing its diarization core.

speech-to-textdiarizationmultilingualself-hostedvoice-agent
◆Current state
Deepgram shipped a new batch speaker diarization architecture as an opt-in diarize_model parameter, then made it the default in the May self-hosted release. Alongside it, profanity filtering rolled out to 50+ monolingual languages and then to all multilingual models, numerals support reached Russian, Romanian, and Hebrew, and Gemini 3.1 Flash Lite replaced its preview version in the Voice Agent API.
◆Where it's heading
Two tracks run at once. The model track is a genuine architecture replacement in diarization, moved from opt-in to default inside a week and a half, which is a fast promotion for a core speech component. The coverage track is unglamorous breadth — languages, filters, numerals — that determines whether the platform can be adopted outside English-first markets. The Voice Agent work is managed-model plumbing rather than a direction of its own.
◆Prediction
Diarization v2 should reach streaming after landing in batch and self-hosted, since that is where the earlier architecture is weakest for live agents. Expect the language-coverage releases to continue at the same steady cadence.

◆Recent moves

  1. 2mo ago

    Profanity Filtering Now Supported for All Multilingual Models; Korean Spacing Improvements

    Profanity filtering extends from monolingual models to every multilingual model, with Korean spacing improvements alongside. The second half of a coverage push that started with 50+ languages the week before — routine breadth work, and the kind that decides non-English deployments.

  2. 2mo ago

    Gemini 3.1 Flash Lite Now Available

    A managed Google model reaches standard tier in the Voice Agent API, replacing its preview version. Menu maintenance rather than a capability change.

  3. 2mo ago

    Numerals Support Now Available for 3 New Languages: Russian, Romanian, and Hebrew (Monolingual Models)

    Numerals support arrives for Russian, Romanian, and Hebrew on monolingual models. Narrow but concrete formatting depth in exactly the markets the parallel profanity-filter rollout targets.

    View source ↗
  4. 3mo ago

    Self-hosted May release ships Diarization v2 by default

    The May self-hosted images promote the new diarization architecture from opt-in to default. Ten days from launch to default is a fast promotion for a core speech model, and it is the strongest evidence the v2 rollout is going well.

    View source ↗
  5. 3mo ago

    Profanity Filtering Now Available in 50+ Languages

    Profanity filtering lands across 50+ monolingual languages, automatically detecting and redacting offensive terms. Compliance-driven breadth that matters for regulated and consumer-facing deployments.

  6. 3mo ago

    Diarization v2: Improved Batch Speaker Diarization

    ⚡ SPARK

    A new batch diarization architecture, opt-in via diarize_model and default in self-hosted images within two weeks. Replacing the model that decides who spoke when is a deeper change than the coverage releases surrounding it.