OmniVoice Voices Five Minutes of Russian Text in 44 Seconds, But Gets the Stress Wrong
The free OmniVoice model reads text aloud and could replace a human voice actor in a voice bot or a training video. Raft engineer tested it on an excerpt from the short story "The Gentleman from San Francisco": five minutes of audio took 44 seconds to generate, about seven times faster than the recording itself. But in places, the ends of phrases get cut off, words are swallowed, and stress goes wrong. Good enough for drafts; a final video needs proofreading.
AI-processed from Raft; edited by Hamidun News
Imagine you need to voice a hundred phrases for a voice bot or a training course, and a week later the texts change. With a live voice actor, that means another studio session, recording, and editing: expensive and slow. Text-to-speech (TTS) models, programs that read text aloud, used to sound robotic, but they have become noticeably more natural.
Raft DS engineer Danil Muzafarov tested the new free model OmniVoice on Russian. The main result: the model generated five minutes of audio from an excerpt of the short story "The Gentleman from San Francisco" in 44 seconds, roughly seven times faster than the recording's length. But the speech still contained artifacts: audio glitches such as a cut-off phrase ending, swallowed words, and incorrect stress.
What OmniVoice can do
The authors claim support for more than 600 languages and three operating modes. The first is voice cloning: you give the model a short voice recording and its transcript, and it then reads any text in that voice. The second is creating a voice from a description. There is a caveat here: a free-form phrase like "low voice of a professional announcer" will not work. The instruct field accepts a limited set of parameters, such as "male, low pitch, russian accent," so the description has to be assembled from ready-made tags. The third mode is automatic voice selection: you pass only the text, and the model chooses the rest.
The author also tested non-verbal tags. You can insert [laughter], [sigh], [surprise-oh], [question-ah], and [dissatisfaction-hnn] into the text, and the speech gains laughter, a sigh, surprise, or displeasure. He ran a business dialogue about funnel conversion and a quarterly report, then checked whether the model returned to a neutral tone after emotional insertions. For voice assistants and training, this matters: monotony is most noticeable in long recordings.
Long narration: fast, but flawed
Long texts are the classic weak point of speech synthesis: artifacts can appear by the middle of a recording, and the timbre can drift. For the test, he used an excerpt from "The Gentleman from San Francisco." The text was split into 15-second chunks (the audio_chunk_duration and audio_chunk_threshold parameters). The result: about five minutes of audio in 44 seconds of generation. For batch voiceover, speed is unlikely to be the bottleneck. However, the material does not say what hardware the computation ran on, so the figure may differ on your GPU.
The author rated the quality as good but with noticeable defects: sentence endings are cut off, words are swallowed, and some stress is wrong. He suspects that tuning hyperparameters (generation settings) could improve this. For now, that is a hypothesis, not a verified result.
How the author checked business scenarios
Pretty examples from the repository say little about real tasks: real texts are full of dates, sums, full names, abbreviations, English terms, and product names. So Danil built his own benchmark with 10 sentences per category. He checked full names, dates and times, numbers, percentages and monetary amounts, abbreviations, homographs (words with different stress, such as "zamok"), and English words inside Russian text. Scoring ran from 0 to 5: zero means unintelligible or missing speech, five means clean pronunciation with no stress or intonation errors.
The available part of the material contains no final scores for these categories, so it is too early to judge those numbers. Moreover, a single person judged the results by ear, and the sample is small. It is a useful guide, but not a rigorous model comparison.
What to do in practice
If you are considering replacing a voice actor, it makes sense to split the tasks. For prototypes of voice bots, internal training materials, and quick voiceover drafts, OmniVoice looks like a workable option: speed is high, voice cloning and emotional tags come out of the box, and the open license gives you control over your infrastructure. For final advertising and customer-facing IVR scripts (automated phone menus), proofreading is still needed: customers will hear the stress errors and clipped phrases right away.
The main advice from this test: don't trust the demos. Instead, assemble your own set of 50–100 real phrases from your company, with customer names, sums, dates, abbreviations, and product names. Run them through the model, listen, and score them on a simple 0-to-5 scale. In one evening, you will know whether the quality is sufficient and how much manual editing to plan into the process.
Key Questions Answered
What OmniVoice can do?
The authors claim support for more than 600 languages and three operating modes. The first is voice cloning: you give the model a short voice recording and its transcript, and it then reads any text in that voice. The second is creating a voice from a description. There is a caveat here: a free-form phrase like "low voice of a professional announcer" will not work.
What are the key details on Long narration: fast, but flawed?
Long texts are the classic weak point of speech synthesis: artifacts can appear by the middle of a recording, and the timbre can drift. For the test, he used an excerpt from "The Gentleman from San Francisco." The text was split into 15-second chunks (the audio_chunk_duration and audio_chunk_threshold parameters). The result: about five minutes of audio in 44 seconds of generation. For batch voiceover, speed is unlikely to be the bottleneck.
How the author checked business scenarios?
Pretty examples from the repository say little about real tasks: real texts are full of dates, sums, full names, abbreviations, English terms, and product names. So Danil built his own benchmark with 10 sentences per category. He checked full names, dates and times, numbers, percentages and monetary amounts, abbreviations, homographs (words with different stress, such as "zamok"), and English words inside Russian text.
Test Frontier LLMs & Autonomous Agents in One Workspace
Run Claude Opus 5.5, GPT-6, Gemini 3.1, and DeepSeek side-by-side in Nomad AI (nomadnet.ai) — unified model routing, deep research, and multimodal workflows under one subscription.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.