← Back to feed

Soniox shipped TTS v2 on August 11. The model is available now as tts-rt-v2 across their API and platform.

The headline number is $0.70 per generated hour. That pricing puts it well below competing frontier TTS offerings, which have generally sat in the $1.50 to $5+ range for equivalent quality tiers. Soniox is making a price-performance argument: comparable output quality, purpose-built streaming architecture, at roughly half to one-fifth the cost.

What Changed From v1

Soniox rebuilt TTS v2 from the ground up. The model architecture, audio codec, and inference engine are all new. The company’s public description emphasises three things that distinguish the new system from both its predecessor and from existing TTS APIs:

Character-level timestamps. Every generated character has a precise timestamp. This is the infrastructure that makes voice agents actually usable in conversational contexts. An agent that knows where in a sentence it is can stop cleanly when interrupted and resume from the right position without repeating itself. Without character-level timestamps, voice agents either barrel through interruptions or restart awkwardly. This is a design-level commitment to agent use, not a bolted-on feature.

Audio tags. Developers can insert control signals directly into the text being synthesised: whispering, excitement, laughter, pauses, and other delivery changes. The voice performance becomes programmable at the text level without separate model calls or post-processing. For applications that need expressive spoken output — customer service, interactive entertainment, voice interfaces — this changes what is possible in a single API call.

Multilingual mixing. The model handles over 60 languages and switches between them within a single utterance without requiring language specification or separate model routing. Natural language mixing — a question in English, a name in Japanese, a phrase in French — is handled in one pass.

Architecture for Agents

The agent use case is explicit in the design. Low-latency streaming lets an agent begin speaking before the full response is generated. The character-level timestamps enable clean interruption: when a user speaks over the agent, the system knows exactly where the utterance was cut off and can resume from that point. The result is a voice interaction that does not sound like a queue-based text-to-speech system spliced onto a language model.

Vapi, which builds voice agent infrastructure, selected Soniox as its speech-to-text provider for several of its voice agent presets. That relationship suggests the Soniox system has been validated in production voice agent pipelines, not just benchmarked in isolation.

The Pricing Argument

The most directly competitive data point is Grok Voice APIs, which xAI priced at $0.10/hr for transcription but $4.20/M characters for synthesis. Google’s Gemini 3.1 Flash TTS competes on language breadth but not on per-hour pricing. OpenAI’s GPT-based speech systems are integrated into broader platform pricing structures.

At $0.70/generated hour, Soniox is targeting the layer of the voice stack where cost-per-interaction matters: high-throughput customer service agents, always-on voice assistants, and applications that cannot absorb $3–5/hr TTS costs at scale.

Voice cloning is included in the model. An application that needs consistent brand voices or personalised agent personas does not need a separate fine-tuning workflow or a separate product tier.

What It Does Not Address

Soniox is a specialist provider, not a full-stack voice platform. TTS v2 handles synthesis. Developers still need STT (Soniox has its own), an LLM, and orchestration to build a complete voice agent. The value proposition is that the TTS layer no longer needs to be the bottleneck — either on latency, on cost, or on language coverage.

The comparison to be made at this price point is not with frontier labs’ bundled offerings, but with the prior generation of TTS APIs that voice agent builders have been routing around because of quality or cost constraints. That is the gap Soniox is positioning into.