GradiumBENCHMARKS

Read on November 19, 2026

Gradium products

Text-to-Speech (Gradium TTS)

Real-time streaming text-to-speech with expressive voices in five languages over a single streaming API.

· Text-to-Speech under 50ms that nails the hard parts

· 49 ms time to first audio (P50), Speko TTS benchmark

· 50 ms TTFA P50, 0 ms leading silence, 5 ms P25–P75 spread (perceived TTFA, Coval benchmark)

· 81.0% hard-case pass rate: 500 sentences, 5 languages

Source : gradium.ai/text-to-speech

Speech-to-Text (Gradium STT)

Streaming transcription built for live conversation, with semantic turn detection, keyword boosting and tunable latency over one WebSocket.

· Semantic turn detection signal every 80ms

· Tunable delay from 0.56s to 3.84s

· Keyword boosting / custom pronunciation: up to 500 words, no retraining; each word's odds can rise ~20x

· Once a word appears, it stays (no rewriting of earlier words)

Source : gradium.ai/speech-to-text

Voice Cloning (Instant and Pro)

Clone a voice from a 10-second sample (Instant) or fine-tune a Pro Voice Clone on 30 minutes to a few hours of audio.

· Clone any voice from 10s of audio; no training step

· Instant clone returns a usable voice in about 3 seconds

· Highest Elo vs ElevenLabs Flash 2.5 in English, French, Spanish and German, in 3,220 blind human comparisons

· 1,000 instant clones a month on every paid plan; unlimited on Enterprise

Source : gradium.ai/voice-cloning

Voice Design

Generate a brand-new voice from a short text prompt (age, gender, accent, texture, mood), audition candidates and keep one as a permanent voice ID.

· 4.06 / 5 accent adherence, best of four voice design systems tested

· 83.4% prompt adherence on the InstructTTSEval English benchmark

· 100+ accents

· Voice Design is free and comes with every plan

Source : gradium.ai/voice-design

Live Translation (Gradium Translate: stt-translate and s2s-translate)

Real-time speech translation between five languages in any direction, as text (stt-translate) or speech in a chosen or cloned voice (s2s-translate).

· 20 language pairs: any of five languages into any other

· 3.0 s average end-to-end latency across all language pairs

· Output in any of 1000+ catalogue voices or a clone of your own

· Source language detected automatically

Source : gradium.ai/live-translation

Phonon (on-device TTS)

Gradium's open-weight on-device text-to-speech, running offline on CPUs, phones, small boards and in the browser.

· Three models: Phonon 31m (64 MB, English, Apache 2.0), Phonon 40m (74.6 MB, 5 languages one model per language, CC BY-NC), Phonon 40m multilingual (74.6 MB, 5 languages in one model, CC BY-NC)

· 1.7% WER (Phonon 31m, English) on 500 hard sentences: phone numbers, email addresses, postal codes

· Starts speaking in 7 ms on a MacBook Air M5 CPU; Phonon 31m starts in 38 ms and runs 3.7x faster than real time on a Raspberry Pi 5

· Up to 22x faster than real time on an iPhone 16 Pro

Source : gradium.ai/phonon

Gradbot

Gradium's open-source voice agent framework (Rust orchestration core) coordinating STT, any OpenAI-compatible LLM and TTS in real time.

· Build real time voice agents in about 50 lines of code

· Handles VAD, turn-taking, fillers, interruptions; MCP integration; remote server mode

· Built for fast prototyping, demos, and experimentation; for production use Gradium APIs directly or mature integrations such as Pipecat

· Install: pip install gradbot

Source : gradium.ai/gradbot

Speech-to-Speech (S2S)

Single duplex WebSocket that transcribes, translates and re-synthesizes audio; the only supported S2S model is live translation (s2s-translate).

· Returns both synthesized output audio and translated transcript as they're produced

Source : docs.gradium.ai/guides/speech-to-speech-overview.md

Voice Enhance and Voice Localization

Enhance cleans up an existing voice without changing language or speaker; Localization adapts an existing voice to another language or accent.

· Voice Enhance beta September 24, 2026; out of beta October 8, 2026

· Voice Localization API October 8, 2026; Voice localizer in Studio October 16, 2026

Source : docs.gradium.ai/guides/release-notes/2026-10.md

Gradium Studio

Voice playground and account management at studio.gradium.ai.

· Unlimited seats in the studio on every plan

Source : gradium.ai/llms.txt

TTS models

ModelAPI
Gradium TTS (current default)defaultDefault TTS model; 'gradium-tts-beta' also routes to it until the next beta. Became default September 30, 2026 with more expressive speech and lower latency; cfg_coef deprecated. · around 50ms for the first audio chunk (blog, Oct 1, 2026); under 50 ms at beta launch vs 225 ms for the then-production model (release notes, Sep 22, 2026); 47 ms P50 on Coval (7-day window, read September 24, 2026)
gradium-tts-beta (alias)gradium-tts-betaAlias that follows TTS beta models; currently routes to the default model.
Gradium TTS (August 31, 2026 release)default (at the time); LiveKit Inference id gradium/defaultReads phone numbers, email addresses, IBANs and reference codes with no pre-processing; public beta from July 30, 2026. · TTFA P50 216 ms on Coval, 170 ms faster than the previous model; 30 ms p75 - p25 spread
gradium-tts-202602 (previous generation)gradium-tts-202602Previous model kept temporarily after the June 10, 2026 upgrade; to be gradually phased out.
Phonon 31mphonon-31m-en (Hugging Face)On-device, 31m parameters, 64 MB, Apache 2.0. · 38 ms to start on a Raspberry Pi 5
Phonon 40me.g. phonon-40m-fr (Hugging Face)On-device, 40m parameters, 74.6 MB, one model per language, CC BY-NC 4.0.
Phonon 40m multilingualOn-device, 40m parameters, 74.6 MB, five languages in a single model, CC BY-NC 4.0.

Integrations

LiveKit Agents

Gradium plugin (livekit.plugins.gradium) for STT, TTS or both in one AgentSession; Gradium TTS on LiveKit Inference as gradium/default

Source : docs.gradium.ai/integrations/agent-frameworks/livekit.md · gradium.ai/llms.txt

Pipecat

GradiumTTSService streams Gradium TTS over WebSocket in a Pipecat pipeline

Source : docs.gradium.ai/integrations/agent-frameworks/pipecat.md

Vapi

Not a built-in provider; connected via custom-voice webhook and custom-transcriber WebSocket

Source : docs.gradium.ai/integrations/agent-frameworks/vapi.md

OpenClaw

Bundled Gradium speech provider for messaging-channel agents (Opus voice notes, 8 kHz u-law telephony)

Source : docs.gradium.ai/integrations/agent-frameworks/openclaw.md

Gradbot

Gradium's own open-source voice agent framework

Source : docs.gradium.ai/integrations/agent-frameworks/gradbot.md

Keenable

Realtime web search for Gradium voice agents (Gradbot demo)

Source : docs.gradium.ai/integrations/web-search/keenable.md

Linkup

Web search API used in a Gradbot hotel-booking voice agent demo

Source : docs.gradium.ai/integrations/web-search/linkup.md

Tavily

Real-time web search for a Gradium voice agent (Paris rental demo)

Source : docs.gradium.ai/integrations/web-search/tavily.md

AWS Marketplace / Amazon SageMaker

Managed SaaS subscription via AWS Marketplace and a deployable model image via SageMaker for in-VPC inference

Source : gradium.ai/llms.txt · gradium.ai/text-to-speech

Baseten

Dedicated Gradium TTS deployment hosted on Baseten, same WebSocket protocol

Source : docs.gradium.ai/guides/self-hosted/baseten-tts.md

Eden AI

Gradium TTS, STT and speech translation available through Eden AI's API, pinned to the EU region

Source : gradium.ai/llms.txt

AudioStack

Gradium voices live on AudioStack's audio production platform

Source : gradium.ai/llms.txt

Yampa

Gradium voice models run natively inside Yampa (customer care agents), SaaS on EU infrastructure or on-premise

Source : gradium.ai/llms.txt

Security and compliance

SOC 2 Type II attestation (completed October 12, 2026; A-LIGN badge)ISO 27001:2022 certification, independently audited by A-LIGN (September 16, 2026)HDS v2 certified (French health data hosting) for hosting the application platform and operating the information system; HDS orgs get strict EU residency and Zero Data RetentionData residency: sessions pinned to the EU or the US (paid plans)Zero Data Retention, self-serve on paid plansGDPR: DPA applies automatically on self-serve plans; voice cloning not treated as Article 9 special category processingCustomer content not used to train modelsTrust Center at trust.gradium.ai