← Back to all sparks
S

Speechmatics

COMMS
Velocity5.0

Speechmatics rolls its Enhanced English model across the stack, citing 89% WER gains on spellouts.

speech-to-textvoice agentsmultilingualmedicalaccuracyenterprise
◆Current state
Speechmatics is a speech-recognition platform whose last quarter has been a coordinated rollout of its Enhanced Operating Point English model from containers through realtime and batch SaaS. The accuracy story is unusually concrete: 69% relative WER improvement on numbers, 89% on spellouts, 42% on mixed alphanumerics. Alongside the model work, the platform is adding voice-agent ergonomics — End of Utterance detection, prefer_current_speaker, speaker sensitivity — and broadening bilingual coverage.
◆Where it's heading
Two threads are running in parallel. Vertical depth: domain-specific models starting with medical, plus a growing list of bilingual code-switching pairs (Tagalog, Malay/English, Tamil/English, Mandarin/English, Arabic/English). Horizontal coverage: each model lands in containers first, then realtime SaaS, then batch SaaS, then appliance — containers function as the proving ground and SaaS as the broad rollout vehicle. The release notes also hint at voice agents being the primary use case Speechmatics is optimising for.
◆Prediction
Expect more vertical-domain Enhanced models beyond medical (legal and finance are the obvious next targets) and a tighter packaging of the voice-agent primitives — End of Utterance, current-speaker locking, low-latency operating points — into something explicitly marketed as a voice-agent SDK or recipe.

◆Recent moves

  1. 5mo ago

    15.7.0 - Containers

    ⚡ SPARK

    The 15.7.0 container release combines the Enhanced English model rollout with HTTP Batch Transcription on persistent workers — the latter is the architectural piece that makes high-throughput voice-agent and contact-center workloads economical on the same infrastructure.

    View source ↗
  2. 5mo ago

    2026.04.23 Realtime SaaS

    Brings the Enhanced English model to Realtime SaaS, including the low-latency ForceEndOfUtterance accuracy improvements that voice-agent builders care about most.

    View source ↗
  3. 5mo ago

    2026.04.16 Batch SaaS

    Same Enhanced English model lands in Batch SaaS — completing the rollout sequence from containers through both SaaS surfaces.

    View source ↗
  4. 6mo ago

    2026.03.12 - Realtime SaaS

    The first Realtime SaaS appearance of the Enhanced English model with the headline alphanumeric accuracy gains — the table of relative WER improvements (69% on numbers, 89% on spellouts) is the version of the story Speechmatics keeps repeating because customers are asked to validate it themselves.

    View source ↗
  5. 6mo ago

    2026.03.09 - Batch SaaS

    Custom-dictionary throughput improvement for short audio plus the Enhanced English model in Batch SaaS — the dictionary speedup quietly lowers the cost of shipping vocabulary-heavy verticals like medical or legal.

    View source ↗
  6. 6mo ago

    1.2.0 Realtime Kubernetes — Redis repopulation and CVE patches

    Realtime Kubernetes 1.2.0 is platform plumbing — Redis repopulation, custom inference recipe cost generation, container version bump, CVE patches. Operationally relevant for SREs running self-hosted, but no user-visible capability change.

    View source ↗