MWS-RESON-ASR-OSD

🇷🇺 Русскоязычное описание ниже / Russian summary below.

RESON ASR OSD is a Russian-language benchmark for automatic speech recognition in the telephone channel. The release contains 12,114 recordings and 37 hours 52 minutes of 8 kHz audio. It is an evaluation set: every configuration has a single test split.

The benchmark has two domains, each in two acoustic conditions:

• 🗣️ General — conversational telephone speech: support requests, automated notifications, and everyday calls. The speech keeps natural disfluencies: pauses, fillers, repetitions, self-corrections, and truncated words. Named entities are a deliberate focus: companies, banks, mobile operators, digital services, devices, and applications.
• 🔢 Numeric — numbers and numeric sequences: phone numbers, personal accounts, bank card numbers, dates, amounts, contract numbers, document series and numbers, INN, and SNILS. All numeric sequences are fictional. Any match with a real document is accidental.

🔇 Quietly is a relatively clean telephone channel, not a studio recording. 📢 Noisily contains background speech as well as noise: television, overlapping voices, and similar interference.

🔗 Official repository: github.com/mts-ai/RESON
🏢 Organization: MTSAIR on Hugging Face
📰 Article on Habr (in Russian): comming soon

🎯 What it measures

1. Recognition of conversational Russian in the telephone channel.
2. Robustness to acoustic noise and background speech.
3. Recognition of named entities and English loanwords.
4. Recognition of numbers and numeric sequences.

⚠️ Acoustic conditions are not paired

Quietly and noisily are thematically comparable slices. They are not paired versions of the same recordings. Compare them as two conditions. Do not subtract one score from the other and report the difference as the effect of noise.

📊 Statistics

Audio is mono, 16-bit PCM, 8 kHz. Durations below are sums of the duration field.

Config	Domain	Condition	Duration	Recordings	Male	Female
general_quietly	General	quietly	13:42:44	4,481	2,024	2,457
general_noisily	General	noisily	14:01:05	4,640	2,193	2,447
numeric_quietly	Numeric	quietly	05:01:01	1,463	747	716
numeric_noisily	Numeric	noisily	05:07:27	1,530	736	794
Total			37:52:17	12,114	5,700	6,414

Utterance length is about 11 seconds in General (from 6.0 to 25.2 s) and about 12 seconds in Numeric (from 6.0 to 28.7 s).

📝 Reference transcriptions

Every recording has two references.

• text — spoken form, with punctuation and capitalization. Numbers and loanwords are written mostly as they are pronounced, in Cyrillic.
• e2e_text — written form, with punctuation and capitalization. Numbers are digits. Loanwords and brand names that are conventionally Latin are written in Latin.

Fillers such as «аа» and «ээ» are kept in both fields. In Numeric, text and e2e_text differ on every row. In General they differ where a name, a loanword, or a number changes orthography.

	text	e2e_text
Named entity	Я пополнил Теле два через Райффайзен, а денег нет.	Я пополнил Теле2 через Райффайзен, а денег нет.
Loanword	Айфон показывает сеть, аа но Авито не грузится.	IPhone показывает сеть, аа но Avito не грузится.
Number in General	Алло. Билайн. Почему скорость упала до ста килобит?	Алло. Билайн. Почему скорость упала до 100 килобит?
Numeric sequence	Мой СНИЛС шестьсот шестьдесят восемь пятьсот пятьдесят шесть пятьсот двадцать шесть семь семь.	Мой СНИЛС 668 556 526 77.

💻 How to load

from datasets import load_dataset

general_clean = load_dataset("MTSAIR/MWS-RESON-ASR-OSD", "general_quietly", split="test")
numbers_noisy = load_dataset("MTSAIR/MWS-RESON-ASR-OSD", "numeric_noisily", split="test")

row = general_clean[0]
audio = row["audio"]          # 1-D array, sampling_rate == 8000
written = row["e2e_text"]     # primary reference
spoken = row["text"]          # spoken-form reference
gender = row["gender"]        # "male" or "female"
duration = row["duration"]    # duration in seconds

Keep the native 8 kHz rate for telephone-channel evaluation. Resample only if the model requires another rate, and state that choice with the score.

📁 Repository Structure

MWS-RESON-ASR-OSD/
├── README.md
├── general_quietly/
│   ├── metadata.jsonl
│   └── wavs/*.wav
├── general_noisily/
│   ├── metadata.jsonl
│   └── wavs/*.wav
├── numeric_quietly/
│   ├── metadata.jsonl
│   └── wavs/*.wav
└── numeric_noisily/
    ├── metadata.jsonl
    └── wavs/*.wav

📋 Data fields

Each line of metadata.jsonl is one recording. file_name is relative to the directory that contains that metadata file.

{
  "file_name": "wavs/51_76326400815_185_196.wav",
  "id": "51_76326400815_185_196",
  "text": "Я пополнил Теле два через Райффайзен, а денег нет.",
  "e2e_text": "Я пополнил Теле2 через Райффайзен, а денег нет.",
  "duration": 10.166375,
  "gender": "male"
}

Field	Type	Description
file_name	string	Path to the wav, relative to metadata.jsonl. Loaded as the audio column.
id	string	Recording id, equal to the wav file name without extension.
text	string	Spoken-form reference.
e2e_text	string	Written-form reference.
duration	float	Duration in seconds.
gender	string	Annotated gender of the speaker on this recording: male or female.

📏 Evaluation protocol

Normalize the hypothesis and the reference the same way before scoring:

1. Lowercase.
2. Map ё to е
3. Remove punctuation.
4. Collapse whitespace.
5. Do not convert digits to words or words to digits.
6. tokens "ээ", "аа" and "мм" have been removed

A system that writes «теле два» where the written reference has «Теле2» is wrong on e2e_text. That difference is part of the benchmark.

📊 Open-source baselines

Public ASR systems were run on this benchmark without fine-tuning on it. The table is WER in percent. Lower is better. These scores were computed with RESON and its default metrics provider. Hypotheses were normalized with the evaluation protocol above, then compared with the spoken-form reference in text. The table is WER against text, not against e2e_text. Average is the unweighted mean of the four subset WERs. A dash means that subset was not part of the run, so Average is left empty.

Vendor	Model	General quietly	General noisily	Numeric quietly	Numeric noisily	Average
Sber	GigaAM v3 rnnt	5.35	18.09	0.71	5.56	7.42
Sber	GigaAM v3 e2e_rnnt*	10.94	21.64	—	—	—
Alpha Cephei	vosk-model-ru	8.40	26.58	1.06	8.24	11.07
Nvidia	parakeet-tdt-0.6b-v3	11.01	33.11	3.60	15.60	15.83
Nvidia	nemotron-3.5-asr-streaming-0.6b	19.39	46.11	8.23	28.37	25.52
OpenAI	Whisper Large v3 Turbo*	12.47	62.15	—	—	—

---

📄 License

MIT License
© 2026 MTS AI

See LICENSE for details.

---

📚 Citation

If you use this dataset in your research, please cite:

@misc{mws_reson_asr_osd_2026,
  title={MWS RESON ASR OSD: A Russian Telephone-Channel ASR Benchmark},
  author={MTS AI Research},
  organization={MTSAIR},
  year={2026},
  url={https://huggingface.co/datasets/MTSAIR/mws-reson-asr-osd},
}

---

🤝 Contacts

• Team: MTSAIR Research
• Email: a.shevchenko@mts.ai

---

🇷🇺 Краткое описание

MWS RESON ASR OSD — открытый бенчмарк MWS AI, разработанный командой Audiogram, для оценки качества русскоязычных ASR-моделей в телефонном канале. Бенчмарк состоит из двух датасетов (General и Numeric), каждый представлен в двух акустических версиях — условно чистой (quietly) и шумной (noisily), содержащие не просто шумы, но и фоновую речь, что позволяет проверить устойчивость модели к звукам из телевизора, посторонним голосам и так далее. Все аудио — 8 кГц. В суммарности датасет содержит 12 114 записей, общая длительность — 37 часов 52 минуты.

• General — русскоязычный датасет телефонных реплик, обращений в поддержку, автоматических уведомлений и бытовых звонков. Для него характерны естественные особенности разговорной речи: паузы, междометия, повторы, оговорки, самокоррекции и отдельные обрывы слов. Особое внимание уделено распознаванию имен собственных: названий компаний, банков, операторов связи, цифровых сервисов, устройств и приложений.
• Numeric — специализированный набор на числа и числовые последовательности: телефонные номера, лицевые счета, номера банковских карт, даты, денежные суммы и платежи, номера договоров, серии и номера документов, ИНН и СНИЛС. Все числовые последовательности вымышлены, любые совпадения случайны.

Для каждой записи в бенчмарке есть два варианта эталонной разметки:

• text — базовая транскрипция с пунктуацией и капитализацией, где числа и заимствованные названия / англицизмы записаны преимущественно в произносительной форме прописью на кириллице;
• e2e_text — нормализованная письменная форма: пунктуация, капитализация, числа записаны цифрами, а англицизмы латиницей.

Дополнительно присутствует информация о поле диктора (gender): male или female.

---

Made with ❤️ by MTS AI Research Team