Introducing OhanTTS: Armenian Text-to-Speech

Community Article
Published November 20, 2026

OhanTTS-banner

We’re excited to introduce OhanTTS, an Armenian-focused text-to-speech (TTS) model for generating natural speech! Alongside Armenian, it generates speech in English, Spanish, Arabic, Russian, Mandarin Chinese, Japanese, and French. Our first release, OhanTTS-v0.1, builds on a multilingual speech model and was further trained on approximately 160 hours of reference recordings across these eight languages, including 61 hours of Armenian speech.

OhanTTS supports reference voice cloning: provide a 3–10 second audio sample and its transcript, and the model generates speech in that voice.

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

Why we built it

We wanted Armenian speech that starts playing right away. Most TTS models generate the whole utterance before returning any audio, so latency grows with input length. OhanTTS instead generates audio incrementally and streams it as it is produced, so you hear the first words in about a third of a second.

OhanTTS streaming pipeline: inputs, frame-by-frame generation, waveform decoding, and streamed audio

You still send the full text upfront; only the audio is streamed.

How Well Does It Perform?

Synthesis quality

We evaluated OhanTTS on a held-out set of Armenian synthesis requests with unseen speakers and texts.

Metric Result
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 requested text. SIM-o measures how closely the generated voice matches the reference speaker.

Streaming performance

We benchmarked the served model over local HTTP on a single shared NVIDIA GB10, with streamed PCM output:

Concurrent clients Time to first audio ↓ Real-time factor ↓ Requests/s ↑ Audio seconds/s ↑
1 331 ms 0.178 0.69 5.59
2 464 ms 0.236 0.90 7.30
4 651 ms 0.302 1.36 11.02
8 1,142 ms 0.505 1.88 15.33
16 2,169 ms 0.841 2.52 20.52

Time to first audio and throughput under concurrent load

With a single client, the first audio arrives in 331 ms, and even with 16 concurrent clients every stream is generated faster than real time.

Performance depends on hardware, request length, and server load.

Listen to OhanTTS in Action

Armenian

Same text, different 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

OhanTTS can also create a voice from a short description, without any reference recording:

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

OhanTTS also speaks English:

"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."

Available for Everyone

We're releasing OhanTTS under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license, making it freely available for research, educational, and other non-commercial use.

  • OhanTTS-v0.1: streaming Armenian-focused text-to-speech with voice cloning.

OhanTTS is served through SGLang-Omni with an OpenAI-compatible /v1/audio/speech endpoint. The cookbook includes complete voice-cloning and streaming examples.

Known Issues

OhanTTS is still an early release, and there are some known limitations:

  • Long utterances: Each request generates up to approximately 20 seconds of audio; longer inputs should be split.
  • Armenian punctuation: Marks such as «՝», «՛», «՜» and «՞» are not always reflected correctly in pauses, stress, and intonation.
  • Unfamiliar names: Rare names and foreign words may be mispronounced.
  • Code-switching: Switching between languages within an utterance can occasionally cause pronunciation or generation errors.
  • Reference quality: Results depend on a clean reference recording and an exact transcript.

Please use reference voices only with the speaker's permission.

What's Next?

This is just the first release. Our next major step is bistreaming: accepting text in chunks and starting to speak while the input is still arriving. For applications connected to a language model, this means speech can begin from the first words of a response while the rest is still being written.

We'll also continue working on more robust pronunciation across languages and longer generation.

We're excited to see what the community builds with OhanTTS and welcome feedback.

Acknowledgements

We acknowledge the EuroHPC Joint Undertaking for awarding this project access to the EuroHPC supercomputer LUMI, hosted by CSC (Finland) and the LUMI consortium. OhanTTS-v0.1 was trained on LUMI.

Give OhanTTS a try, and help us make Armenian speech technology better for everyone!

Community

Sign up or log in to comment