Last updated: May 28, 2026
Inworld AI Realtime TTS and Google Gemini Flash TTS are engineered for different categories of voice work. Realtime TTS-2 (research preview) is the #1 realtime TTS on the
Artificial Analysis Realtime TTS Arena, purpose-built for sub-second voice agents, AI companions, and live conversation. Gemini Flash TTS is a high-quality batch generation model from Google with broad multilingual coverage and tight integration into the Google Cloud and Vertex AI ecosystem. The right choice depends on what you are optimizing for: realtime voice loops or offline audio production.
This page is not a leaderboard fight. The honest framing is that voice AI has two distinct categories, and choosing the wrong one for your workload costs you either user experience or production quality. Below is how the two compare across the dimensions that actually matter for each category.
What is the difference between realtime TTS and batch TTS?
Realtime TTS is built around one metric: time-to-first-audio. A voice agent only feels responsive when the first audible byte arrives in under a second. Total generation time matters less than how fast the conversation feels.
Batch TTS optimizes for the opposite. Long-form coherence, voice stability across a 30-minute audiobook chapter, prosody on a 12-minute podcast script, and total throughput per dollar are what counts. A 4-second start latency disappears in a content production pipeline. A 4-second start latency kills a voice agent.
Inworld AI Realtime TTS is engineered for the first category. Gemini Flash TTS is engineered for the second. They do not compete on the same benchmark.
How does Inworld Realtime TTS compare to Gemini Flash TTS at a glance?
Where does Gemini Flash TTS actually shine?
Google's voice work is excellent in the categories it targets. Honest acknowledgments before the rest of the page leans into Inworld's realtime carve-out:
- Long-form audio quality. Gemini Flash TTS produces coherent, well-paced output across long scripts. For audiobook generation, narrated explainers, or full podcast episodes, the quality is genuinely strong.
- Native multi-speaker dialogue. Generating a back-and-forth conversation between two voices in a single call is a real product feature that simplifies podcast and dialogue workflows.
- Multilingual coverage. Broad GA language coverage backed by decades of Google research on text-to-audio and speech recognition.
- Google Cloud ecosystem integration. If your data already lives in Cloud Storage, your inference runs on Vertex AI, your queues are Pub/Sub, and your auth is Workload Identity, Gemini Flash TTS lands inside the same security and billing surface with no new contract.
- Research depth. Google has been publishing foundational text-to-audio work for years. The model lineage matters.
If your workload is offline audio production at scale inside Google Cloud, Gemini Flash TTS is a strong default. The rest of this page is about the realtime category, which is a different problem.
Why is realtime TTS a different category?
A realtime voice agent has a hard upper bound on perceived response time. Roughly: under 800ms feels natural, 800ms to 1.5 seconds feels sluggish, anything above feels broken. That budget covers end-of-turn detection, STT finalization, LLM time-to-first-token, TTS time-to-first-audio, and network. TTS gets a slice of that budget, not the whole thing.
Models engineered for batch generation can hit excellent total quality but cannot start emitting audio inside that budget. That is not a flaw of the model. It is a different design target. Realtime TTS-2 is conditioned on prior audio (not just the transcript) so it can begin emitting expressive, conversational audio immediately, and that closed-loop architecture is the reason it sits at the top of the Realtime TTS Arena.
If you are building a voice agent and you select a batch-optimized TTS to save on price or to consolidate vendors, you ship a product that feels broken to the user. The category choice comes first.
What does Inworld bring to the realtime category?
Five things that matter for live voice loops:
#1 realtime TTS. Realtime TTS-2 leads the Artificial Analysis Realtime TTS Arena. Realtime TTS 1 Max sits inside the top of the same category. Both deliver sub-200ms median time-to-first-audio. Realtime TTS 1 Mini delivers ~120ms median when latency is the absolute priority.
Natural-language steering across 8 dimensions (TTS-2). Place a bracketed instruction at the start of the text (for example [say excitedly], [whisper in a hushed style], [in a calm tone], [say with rising tone]) and the model adjusts emotion, articulation, intonation, volume, pitch, range, speed, or vocal style. Non-verbal tags like [laugh], [sigh], and [breathe] work inline. This is research preview today.
Cross-lingual voice identity. The same voice carries across languages mid-utterance. A character can switch from English to Japanese to Spanish without changing identity. For language-learning apps, multilingual companions, and global voice agents, this is a different product than per-language voices that sound like different people.
Model-agnostic Router under one auth. The
Realtime Router routes to 200+ LLMs from OpenAI, Anthropic, Google, Groq, Fireworks, Mistral, and DeepSeek through a single OpenAI-compatible endpoint.
Realtime Inference, the 1P track, runs Inworld-optimized open-source models (Gemma 4, DeepSeek V3.2/V4, MiniMax-M2.5) on Inworld GPUs. No new contract per provider. Fall back across models on a single key.
Realtime API for sub-second voice loops. STT, Router, and TTS connect over a single WebSocket or WebRTC session. Drop-in OpenAI Realtime protocol replacement. Swap any model at any layer.
What does the realtime user evidence look like?
Real workloads on Realtime TTS today:
- Wishroll / Status (AI companions): 1M users in 19 days, with 95% AI cost reduction after migrating to Inworld and 90+ minute session lengths.
- Janitor (character chat): ~600B tokens per day, treating cache-hit-rate as a first-class metric.
- Latitude (AI Game Master): beat OpenAI by a measurable margin in a three-way A/B for live experiences.
- Bible Chat: 2M to 20M characters per week, 85% TTS cost reduction.
- Tolans: consumer companion app shipping on Realtime TTS for daily live interaction.
- Talkpal: language learning grounded in cross-lingual voice identity.
These are realtime, consumer-scale applications where latency, expressiveness, and cost discipline all have to land at the same time. Inworld 1P-optimized inference is engineered for that combination.
How does the developer experience compare?
The two products live in different ecosystems.
Gemini Flash TTS is a first-class citizen inside Google Cloud. If you are running Vertex AI, your inference, storage, queueing, identity, and billing are already aligned. New voice work fits into a known surface. The tradeoff is provider lock-in: switching LLMs means a different API. Switching TTS means a different API.
Inworld Realtime API is provider-agnostic. The Router exposes an OpenAI-compatible Chat Completions surface, so any code written against the OpenAI SDK works by changing the base_url. The Anthropic SDK works against a /anthropic compatibility layer. STT, Router, and TTS share the same Basic-auth API key. You can run Realtime TTS-2 with anthropic/claude-sonnet-4-6 for the LLM and inworld-stt-1 for transcription inside one WebSocket session, and route a fraction of traffic to a different LLM next week without changing the integration.
For teams already deep in Google Cloud where the answer to every new question is "use Vertex AI," Gemini Flash TTS is the natural fit. For teams that want to pick the best model at every layer of the voice stack and stay flexible as the model landscape moves, Inworld is engineered for that.
What about cost at scale?
No price comparison tables here. The honest framing is that consumer-scale voice products live or die on per-minute cost, and Inworld leans into cost discipline through 1P-optimized inference on the LLM and TTS layers. Wishroll cut AI costs 95% after migrating. Bible Chat cut TTS costs 85%. Janitor treats cache-hit-rate as a metric because at 600B tokens per day, every percentage point matters.
For batch audio production inside Google Cloud, Gemini Flash TTS sits inside existing billing and reserved capacity. For realtime consumer apps where margins compound across millions of sessions, Inworld 1P inference is designed for that scale.
See the
Inworld pricing page and Google's published Vertex AI pricing for current rates.
How do voice cloning and voice design compare?
Both ecosystems offer voice cloning, but the workflows differ.
Inworld Realtime TTS:
- Instant voice cloning from 5 to 15 seconds of audio via
POST /voices/v1/voices:clone
- Professional voice cloning from 30+ minutes of audio as a Growth-tier service
- Voice design from natural-language description (TTS-2): generate a new voice from a written character description, no reference audio required
Google Gemini Flash TTS / Chirp 3 HD:
- Instant Custom Voice in the Vertex AI ecosystem (separate Chirp 3 HD line)
- Studio-quality voices generated from short reference samples
If you are cloning real human voices for character work, both products work. If you are generating an original voice from a written description for a fictional character or branded agent, Realtime TTS-2 voice design is engineered for that.
What does the integration code look like for realtime?
A minimum-viable realtime TTS call against Inworld with the recommended inworld-tts-2 model:
Streaming responses return NDJSON. Each line is {"result": {"audioContent": "base64..."}}. Decode base64 per line and pipe to your audio output.
For the full voice loop with STT, LLM, and TTS over a single WebSocket connection, the
Realtime API follows the OpenAI Realtime protocol with an Inworld
providerData extension for STT, TTS, memory, backchannel, and responsiveness controls.
How do you pick between the two?
Two short decision rules:
Pick Google Gemini Flash TTS when:
- You are generating long-form offline audio (podcasts, audiobooks, dubbing, explainer voiceover)
- Total throughput and breadth of GA multilingual coverage matter more than sub-second start latency
- Your stack is already deep inside Google Cloud and Vertex AI
Pick Inworld AI Realtime TTS when:
- You are building a voice agent, AI companion, live tutor, character chat, or any interactive voice product
- Time-to-first-audio under 200ms is the metric your users feel
- You want natural-language steering, cross-lingual voice identity, and a model-agnostic Router that ties STT, LLM, and TTS together under one auth
- You are optimizing for consumer-scale margins where 1P-optimized inference compounds
How do you get started with Inworld Realtime TTS?
- Try the TTS Playground: Hear Realtime TTS-2 (research preview), 1 Max, and 1 Mini with your own text, or clone a voice from a 5-15s sample.
- Read the Realtime TTS documentation: API reference, NDJSON streaming, audio markup tags, and quickstarts.
- Explore the Realtime API: Drop-in OpenAI Realtime protocol with full control over STT, LLM, and TTS at each layer.
- Browse the Router: OpenAI-compatible endpoint routing across 200+ LLMs with Inworld 1P optimized open-source models on the inside track.