Applied AI Research Portfolio Blog Get In Touch

Introducing OhanTTS v0.1

Streaming Armenian text-to-speech with voice cloning — the first words play in about a third of a second.

331 msTime to first audio
8Languages
61 hArmenian speech
3–10 sTo clone a voice
OhanTTS v0.1 — Open Armenian Speech Generation

Overview

Today we're releasing OhanTTS v0.1, an open text-to-speech model built for Armenian. It clones a voice from a few seconds of audio and streams speech as it is generated, so listeners hear the first words almost immediately.

Armenian comes first, but OhanTTS also speaks English, Spanish, Arabic, Russian, Mandarin Chinese, Japanese and French. We started from a multilingual speech model and trained it further on about 160 hours of reference recordings across those eight languages, 61 hours of them in Armenian.

The name comes from Dzenov Ohan, Ohan the Loud-Voiced, from the epic Daredevils of Sassoun. He shares it with OhanASR, our speech recognition model: OhanASR listens and OhanTTS speaks.

Voice cloning

Give OhanTTS a 3–10 second recording and its transcript, and it will speak any new text in that voice. Below, Ohan himself introduces the model: the left clip is the reference, the right one was generated from it.

Reference recording
Ասում են՝ երբ Ձենով Օհանը գոռում էր, ամբողջ Սասունը լսում էր։ Նրա ձայնը հասնում էր սարերից այն կողմ։
They say when Dzenov Ohan shouted, all of Sasun could hear. His voice reached beyond the mountains.
Generated by OhanTTS
Բարև, ես Ձենով Օհանն եմ՝ Սասունից։ Ուրախ եմ, որ այսօր կարող եմ խոսել ձեզ հետ։
Hello, I'm Dzenov Ohan, from Sasun. I'm glad I can talk with you today.

Built for streaming

Most TTS models synthesize the whole utterance before returning any audio, so the longer the text, the longer the silence. That's fine for audiobooks and too slow for a voice agent.

OhanTTS predicts audio one frame at a time and decodes each block to a waveform as soon as it is ready, so the first words reach the listener in about a third of a second while the rest is still being generated. The text is still sent in one request. Only the audio is streamed.

OhanTTS streaming pipeline: reference audio and text go into a causal Transformer that predicts one frame of 8 codec IDs at a time, fed back until speech ends; a waveform decoder turns blocks of frames into audio that is streamed out
Figure 1. The streaming pipeline. A causal Transformer predicts one frame of codec tokens at a time, and the waveform decoder turns each block into audio straight away.

Results

Synthesis quality

We tested OhanTTS on held-out Armenian synthesis requests, with speakers and texts it had not seen in training.

MetricResult
Word error rate (WER) ↓14.70%
Character error rate (CER) ↓6.15%
Speaker similarity (SIM-o) ↑0.693

WER and CER compare an automatic transcription of the generated speech with the text that was requested, so they measure intelligibility. SIM-o measures how close the generated voice is to the reference speaker.

Streaming performance

We served the model over local HTTP on a single shared NVIDIA GB10, streaming PCM audio, and increased the number of concurrent clients. With one client, audio starts in 331 ms. Even with 16 clients, every stream is still generated faster than real time (real-time factor below 1).

Figure 2. Latency and speed under concurrent load. A real-time factor below 1× means audio is generated faster than it plays.
Concurrent clientsTime to first audio ↓Real-time factor ↓Requests/s ↑Audio seconds/s ↑
1331 ms0.1780.695.59
2464 ms0.2360.907.30
4651 ms0.3021.3611.02
81,142 ms0.5051.8815.33
162,169 ms0.8412.5220.52

Measured on a single shared NVIDIA GB10. Results depend on hardware, request length and server load.

Listen

One sentence, two voices

Ամեն անգամ, երբ վերադառնում եմ Երևան, առաջինը գնում եմ Կասկադ՝ քաղաքին վերևից նայելու։
Every time I come back to Yerevan, the first thing I do is go to Cascade to look at the city from above.
Female voice
Male voice

Voice design

No reference recording? OhanTTS can also build a voice from a short text description.

Շնորհակալություն զանգահարելու համար։ Ձեր զանգը մեզ համար կարևոր է և շուտով կպատասխանի մեր առաջին ազատ մասնագետը։
Thank you for calling. Your call is important to us and our first available specialist will answer shortly.
Voice description male, moderate pitch

English, too

"Hi, I'm Ohan. I was built for Armenian, but I speak English too, and I'd be happy to read anything you'd like."

Get the model

OhanTTS is released under CC BY-NC 4.0, free for research, education and other non-commercial use. For commercial licensing, custom voices or on-prem deployment, get in touch.

It is served through SGLang-Omni, which exposes an OpenAI-compatible /v1/audio/speech endpoint, so most existing TTS clients work with it unchanged. The cookbook has complete voice-cloning and streaming examples.

Known limitations

This is an early release. Here's what we know isn't right yet:

  • Long inputs. Each request produces up to about 20 seconds of audio, so split longer text into several requests.
  • Armenian punctuation. Marks like «՝», «՛», «՜» and «՞» don't always come through in pauses, stress and intonation.
  • Unfamiliar names. Rare names and foreign words can be mispronounced.
  • Code-switching. Switching languages mid-utterance occasionally causes pronunciation or generation errors.
  • Reference quality. Cloning works best with a clean recording and an exact transcript.
Clone responsibly. Only use someone's voice as a reference with their permission.

What's next

The next big step is bistreaming: accepting text in chunks and speaking while it's still arriving. Connected to a language model like Ayb, speech could start from the first words of a reply while the rest is still being written. We're also working on more robust pronunciation across languages and longer generation.

Try it and tell us what you build, and what breaks. Feedback is how v0.2 gets better.