A speech AI company founded in 2020 in San Mateo, California, led by CEO Klemen Simonic, with roughly 18 employees and backing from Samsung Venture Investment. Its core is a unified speech-to-text model covering 60+ languages that supports mid-sentence language switching without changing models or restarting the stream. Real-time streaming runs over a WebSocket API, returning partial tokens in about 100ms with built-in endpoint detection for voice-agent turn-taking, while an asynchronous REST API handles file transcription. It also offers translation, text-to-speech, and voice cloning. Compliance claims cover SOC 2 Type 2, ISO/IEC 27001:2022, HIPAA, and GDPR, with audio processed only in memory and not stored, and regional deployment in the US, EU, Japan, and India. Pricing is usage-based: real-time STT at roughly $0.12/hour and asynchronous or media transcription from about $0.10/hour, with free credits for evaluation. In June 2026, Soniox’s STT API was integrated directly into the Tencent Cloud RTC console across 200+ countries. Its accuracy claims come from an in-house March 2025 benchmark covering 60 languages against nine providers using double-verified human ground truth on real YouTube audio—thorough in method, but still vendor-run.
Pros & Cons:
Pros:
A single model covering 60+ languages with mid-sentence language switching removes the need to route callers to different models by language; real-time WebSocket streaming with ~100ms partial tokens and built-in endpoint detection is directly useful for voice-agent turn-taking; regional deployment in the US, EU, Japan, and India addresses data-residency requirements that many STT providers ignore; compliance claims cover SOC 2 Type 2, ISO 27001, HIPAA, and GDPR, with audio processed in memory and not stored; pricing at $0.10–$0.12/hour is low relative to many managed STT services; the June 2026 Tencent Cloud RTC integration is a verifiable enterprise partnership rather than a logo claim.
Cons: Accuracy claims are vendor-run—the March 2025 benchmark covers 60 languages against nine providers with a documented method, but no independent third-party comparison (like Epoch AI’s work on detectors) surfaced to confirm the “significantly leading” result; growth signals conflict—Crustdata shows headcount up 38% year over year while site traffic fell 24% quarter over quarter, with a growth score of 60/100; pricing across real-time, asynchronous, translation, and TTS is billed separately, so total cost at volume needs modelling rather than a single per-hour figure; audio is sent to Soniox servers (TLS-encrypted) despite the no-storage policy, so sensitive recordings that must stay on-device require a local engine like Whisper instead; funding details beyond Samsung Venture Investment are not disclosed; name collision with Sonix, an unrelated Swiss game-communications company, requires disambiguation.
Who is it for?:Voice-agent and real-time transcription builders that need low-latency streaming STT with endpoint detection; products serving multilingual users where mid-sentence language switching matters; teams that require EU, Japan, or India data residency for speech processing; developers who want usage-based pricing from $0.10/hour rather than seat or platform contracts.