Skip to main content
Back

Inworld's Realtime TTS-2 Beats Cartesia at 58% Lower Cost

Inworld's Realtime TTS-2 unseats Cartesia Sonic 3.6 on the Controlled Voice Arena, delivering 100+ language coverage and plain-English delivery instructions at under half the price.

  • Inworld Realtime TTS-2 takes #1 on the Controlled Voice Arena at 1,123 Elo, 4 points ahead of Cartesia Sonic 3.6.
  • Ranks #2 on the Provider Voice Arena at 1,252 Elo, behind Sonic 3.6 and ahead of Qwen-Audio-3.0-TTS-Plus.
  • Priced at $20.83 per 1M characters, undercutting Sonic 3.6 ($49) and Eleven v3 ($100).
  • Supports 100+ languages with voice identity held across mid-generation language switches.
  • Accepts plain-English delivery instructions and non-verbal cues like [laugh], [sigh], [yawn] inline.
  • Available now via Inworld API with OpenAI Realtime protocol compatibility and sub-100ms p99 latency.

Inworld Realtime TTS-2 has taken first place in Artificial Analysis’ Controlled Voice Arena. The general-availability model scored 1,123 Elo, four points above Cartesia Sonic 3.6. Cartesia retains the lead in the Provider Voice Arena and delivers higher raw throughput, leaving developers with a close quality result and clearer trade-offs around voice selection, speed, and price.

The Controlled Voice Arena asks each model to clone the same eight reference voices, split evenly between US and UK speakers. Human judges compare identical sentences and score pronunciation, pauses, intonation, and rhythm. Standardizing the voices reduces the influence of voice casting and exposes differences in the synthesis engines.

Four Elo points separate the leaders

Elo is a relative score derived from pairwise human preferences, so its meaning depends on the competing models and accumulated votes. The narrow gap between Inworld and Cartesia could shift as more ratings arrive.

Controlled Voice Arena results
Rank Model Elo
1 Inworld Realtime TTS-2 1,123
2 Cartesia Sonic 3.6 1,119
3 Cartesia Sonic 3.5 1,096
4 Inworld TTS-2 Flash research preview 1,075
5 ElevenLabs Eleven v3 1,062

The Provider Voice Arena evaluates each vendor’s shipping voices, allowing voice design and selection to affect the result. Cartesia Sonic 3.6 leads that board with 1,282 Elo, followed by Inworld Realtime TTS-2 at 1,252 and Qwen-Audio-3.0-TTS-Plus at 1,241.

Provider Voice Arena leaders
Rank Model Elo
1 Cartesia Sonic 3.6 1,282
2 Inworld Realtime TTS-2 1,252
3 Qwen-Audio-3.0-TTS-Plus 1,241

Inworld’s earlier preview recorded 1,209 Elo on the provider board. The general-availability release improves that score by 43 points.

Quality, throughput, and price

Production voice systems must balance judged quality with generation speed and operating cost. Artificial Analysis reports the following throughput and listed prices for several prominent models:

Text-to-speech performance and pricing
Model Controlled Elo Throughput Price per 1M characters
Inworld Realtime TTS-2 1,123 106 chars/sec $20.83
Cartesia Sonic 3.6 1,119 124 chars/sec $49
Speechify Simba 3.2 Not listed 97 chars/sec $10
Qwen-Audio-3.0-TTS-Plus Not listed Not listed $27.59
ElevenLabs Eleven v3 1,062 40 chars/sec $100

Cartesia processes about 17% more characters per second than Inworld. At the listed rates, Inworld costs about 58% less than Cartesia. Compared with Eleven v3, Inworld provides roughly 2.65 times the throughput at a 79% lower per-character price. Contract terms, volume discounts, and regional infrastructure can change production costs.

Audio context shapes the delivery

Inworld built Realtime TTS-2 for two-way conversation. According to the company’s launch post, its closed-loop architecture feeds audio from earlier conversation turns into the synthesis process. That audio carries pace, stress, hesitation, and other signals that a transcript omits, allowing the generated response to reflect how the user spoke.

Natural-language stage directions provide delivery control through bracketed instructions such as [speak tired but warm, like she just got home from a long day]. The model also supports six bracketed non-verbal cues that render as audio events:

  • [laugh]
  • [breathe]
  • [clear throat]
  • [sigh]
  • [cough]
  • [yawn]

These controls let applications adjust delivery from ordinary text input without constructing detailed SSML documents or mapping every line to a fixed emotion preset.

One voice across languages

Inworld offers 95 ready-made voices covering 14 native languages, with cross-language generation available for more than 100 languages. The company says a selected voice preserves its identity when a script changes languages within one generation, which supports bilingual agents and localized brand voices.

A multilingual request can place several languages in the same text field:

text = "I'll grab a coffee. ¿Quieres uno? お疲れさま。"
# One voice, three languages, one continuous generation

Teams should test accent quality, pronunciation, code-switching, and proper nouns in every target language because aggregate leaderboard scores do not expose language-specific failure rates.

Latency and integration details

Realtime TTS-2 is generally available through the Inworld API and Inworld Realtime API. Inworld also offers TTS-2 Flash for workloads governed by tighter latency, volume, or unit-cost limits. Both models support WebSocket streaming.

Reported latency characteristics
Model Reported result Intended workload
Realtime TTS-2 Under 100 ms at p99 in Inworld’s launch tests Conversational quality and delivery control
TTS-2 Flash 25 ms time to first byte in Coval benchmarking Low-latency and high-volume applications

The reported figures use different metrics and test setups. Production evaluation should measure time to first audio, full-generation latency, network distance, concurrency, audio format, and prompt length under the application’s expected load.

The Realtime API supports the OpenAI Realtime protocol, which can reduce migration work for teams using that event model. Existing Inworld customers moving from version 1.5 need to change the model identifier. Developers should still verify authentication, event schemas, error handling, audio encodings, and interruption behavior before deployment.

An illustrative synthesis payload is:

{
  "text": "I missed you. How was today? [speak warm and soothing]",
  "voiceId": "Craig",
  "modelId": "inworld-tts-2",
  "audioConfig": {
    "audioEncoding": "LINEAR16",
    "sampleRateHertz": 48000
  }
}

Choosing by workload

Realtime TTS-2 targets applications that must respond to live users and adjust delivery from conversational context. Relevant workloads include:

  • Customer-support agents that adapt their delivery after complaints, confusion, or escalation
  • Game characters and companions that maintain a consistent voice across long sessions
  • Multilingual voice-over that preserves one speaker identity across localization
  • Order numbers and confirmation codes that require exact readback through Inworld’s verbatim tag

Cartesia Sonic 3.6 offers higher throughput and leads the Provider Voice Arena, making it a candidate for products centered on a vendor-supplied signature voice. Speechify Simba 3.2 has the lowest listed per-character price in this comparison. Eleven v3 remains an option for teams evaluating long-form narration, although the controlled ranking and throughput figures favor Inworld for the conversational workload measured here.

The four-point controlled lead provides limited separation on judged quality. A production decision should also test interruption handling, tail latency, pronunciation, language consistency, voice availability, rate limits, observability, and total cost under realistic concurrency.