Introducing Teen V1: Expressive multilingual speech at production scale

Teen V1 is a 2.7-billion-parameter text-to-speech model built for expressive multilingual speech, fast synthesis, and zero-shot voice cloning. It uses a pretrained large-language-model backbone, an IPA-centred speech-training stage, and language-group post-training to deliver broad language coverage without requiring one large model to carry every language at once.

This release is built around a practical constraint: multilingual speech should be expressive and production-ready without making inference unnecessarily expensive. Teen V1 reaches that point by concentrating each deployed model on a compact group of related languages and serving those models together as one product.

A 2.7B-parameter speech model

Teen V1 has 2.7 billion parameters and uses an LLM backbone. Starting from a pretrained language model means training does not begin from an empty representation of text. The backbone already contains useful knowledge about written language, context, syntax, and relationships between tokens.

The speech-training budget can therefore focus on turning language representations into speech: pronunciation, timing, prosody, speaker characteristics, and expressive delivery. This is materially more compute-efficient than learning general language representation and speech generation from random initialization in the same training run.

IPA pre-training

Before language-group specialization, Teen V1 is trained on International Phonetic Alphabet representations. IPA provides a shared phonetic notation across languages, allowing the model to learn speech in terms of pronounceable units rather than treating every writing system as an unrelated problem.

This stage gives the model a common pronunciation foundation across different scripts and orthographies. A sound that appears under different spellings can still be represented in a shared phonetic space. The result is a stronger starting point for multilingual speech training, especially when the supported set spans Latin, Arabic, Devanagari, Cyrillic, Greek, and East Asian writing systems.

IPA pre-training does not replace language-specific learning. Pronunciation alone is not enough to model rhythm, stress, phrasing, local phonological rules, or natural expressive delivery. Those properties are learned during the language-group post-training stage.

Language-group post-training

After IPA pre-training, Teen V1 is post-trained on groups of related languages. Each model typically carries four or five languages. Languages with useful phonetic or linguistic overlap are grouped together so they can share model capacity while retaining language-specific pronunciation and delivery.

This grouping keeps each model focused. A single deployment does not need to carry the full multilingual set for every request, and a language does not have to compete equally with every unrelated language for the same inference capacity. The result is a collection of compact multilingual models rather than one monolithic model stretched across the complete catalogue.

From the product surface, the models operate as one network: a request is handled by the appropriate model while the API remains consistent across languages.

This model network is what allows Teen V1 to combine broad multilingual coverage with low serving cost. Capacity is placed where it is useful, related languages share a model, and inference does not pay the full cost of the complete language catalogue on every request.

Supported languages

Teen V1 supports 23 languages: Arabic, Chinese, Dutch, English, French, German, Greek, Gujarati, Hindi, Italian, Japanese, Korean, Marathi, Norwegian, Polish, Portuguese, Punjabi, Rajasthani, Russian, Spanish, Thai, Turkish, and Urdu.

The supported set spans multiple scripts, pronunciation systems, and conventions for word boundaries and text normalization. The public interface remains the same across the catalogue; the language-model network handles the underlying specialization.

Zero-shot voice cloning

Teen V1 supports zero-shot voice cloning from reference audio. A reference recording is supplied at request time, and the target text is synthesized in the cloned voice without a speaker-specific training or fine-tuning job.

The cloning interface is direct: reference audio provides the target voice, and text provides the content to be spoken.

Fast synthesis

Teen V1 exposes streaming and buffered synthesis. Streaming is designed for interactive playback, where the first audio should arrive quickly and generation should continue faster than the audio is consumed. Buffered synthesis is designed for complete-file generation, where total generation speed is the relevant measure.

Real-time factor, or RTF, is complete generation time divided by generated audio duration. An RTF below 1.0 means the system generates audio faster than real time.

Measured performance

A production benchmark completed on 26 August 2026 measured P95 latency and throughput for the Teen V1 fast profile.

Streaming

At P95, the first response byte arrived in under 200 milliseconds. Complete generation remained faster than playback, with a 0.781 RTF.

Buffered

Buffered delivery returns the completed waveform as one result. In the measured workload, Teen V1 generated approximately 45 seconds of audio in 16.23 seconds at P95.

Serving economics

Teen V1 is priced at approximately $5 per million input characters. Zero-shot voice cloning and multilingual synthesis are available through the same serving product.

Across the selected current-generation expressive TTS systems in the comparison, Teen V1 reduces normalized list-price synthesis cost by approximately 66% to 95%.

The low price is a consequence of the training and serving strategy. Language knowledge begins from a pretrained 2.7B-parameter LLM backbone, phonetic learning is shared through IPA pre-training, related languages are post-trained together, and the resulting models are served as a network. Each request uses a focused multilingual model rather than paying the inference cost of one model carrying every supported language.