README
PersonaMix
Controlled bilingual (Kazakh–English) benchmark for target-speaker ASR and target-presence detection on overlapping speech, released with Persona-ASR.
Four speakers (two female, two male) each read 11 scripted sentences in both Kazakh and English. Mixtures span 1–3 interfering speakers and SNRs of −3, 0, +3, +6 dB, under same-language (A) and cross-language (B) enrollment; the cross-language condition enrolls a speaker in one language and transcribes them in the other.
Structure
mixtures/— overlapping-speech mixtures by condition:A_samelang_{1,2,3}int/— same-language enrollment, 1–3 interferersB_crosslang_{1,2,3}int/— cross-language enrollment, 1–3 interferers
experiment_jsons/— evaluation manifests per condition and SNR:*_asr.json— target-speaker ASR (positive trials, with target transcript)*_cls.json— target-presence detection (positive + negative trials)all_snr{-3,0,3,6}dB.json— pooled manifests per SNR
experiments.zip— the whole benchmark as a single archive.ASR_88/andASR_88 - {F1,F2,M1,M2}.csv— the 88 source recordings (4 speakers × 11 sentences × 2 languages) and per-speaker scripts.
Manifest fields (experiment_jsons/*.json)
| Field | Description |
|---|---|
mixture_audio | Overlapping mixture |
enrollment_audio | Target-speaker enrollment utterance |
transcript | Target reference transcript |
sample_type | positive / negative (target present / absent) |
target_speaker | Target speaker id (F1, F2, M1, M2) |
enrollment_lang, mixture_lang | Enrollment vs. mixture language (equal = same-lang, differ = cross-lang) |
snr_db | Target-to-interferer SNR |
enrollment_sentence, target_sentence, interferers | Content and interference metadata |
Speakers
Four speakers, anonymized as F1 and F2 (female) and M1 and M2 (male).
Generation
data_generation/personamix/generate_personamix.py in the Persona-ASR repository regenerates the manifests and mixtures from ASR_88/. The manifests it writes are identical to the released ones, and in our checks the mixtures match the released audio to within one 16-bit sample value. The script's docstring describes the mixing recipe.
License and citation
Released under CC BY 4.0. Please cite Persona-ASR.
@article{meiramov2026personaasr,
author = {Meiramov, Rakhat and Rakhimzhanova, Tomiris and Taibassarov, Adil and Makhataeva, Zhanat and Varol, Huseyin Atakan},
title = {Persona-ASR: Bilingual Target-Speaker Speech Recognition for Kazakh--English Overlapping Speech},
journal = {Machine Learning and Knowledge Extraction},
year = {2026},
volume = {8},
number = {8},
pages = {246},
doi = {10.3390/make8080246}
}
Funding
This research is funded by the Committee of Science of the Ministry of Science and Higher Education of the Republic of Kazakhstan (Grant No. BR24993001).
Similar repos
README
PersonaMix
Controlled bilingual (Kazakh–English) benchmark for target-speaker ASR and target-presence detection on overlapping speech, released with Persona-ASR.
Four speakers (two female, two male) each read 11 scripted sentences in both Kazakh and English. Mixtures span 1–3 interfering speakers and SNRs of −3, 0, +3, +6 dB, under same-language (A) and cross-language (B) enrollment; the cross-language condition enrolls a speaker in one language and transcribes them in the other.
Structure
mixtures/— overlapping-speech mixtures by condition:A_samelang_{1,2,3}int/— same-language enrollment, 1–3 interferersB_crosslang_{1,2,3}int/— cross-language enrollment, 1–3 interferers
experiment_jsons/— evaluation manifests per condition and SNR:*_asr.json— target-speaker ASR (positive trials, with target transcript)*_cls.json— target-presence detection (positive + negative trials)all_snr{-3,0,3,6}dB.json— pooled manifests per SNR
experiments.zip— the whole benchmark as a single archive.ASR_88/andASR_88 - {F1,F2,M1,M2}.csv— the 88 source recordings (4 speakers × 11 sentences × 2 languages) and per-speaker scripts.
Manifest fields (experiment_jsons/*.json)
| Field | Description |
|---|---|
mixture_audio | Overlapping mixture |
enrollment_audio | Target-speaker enrollment utterance |
transcript | Target reference transcript |
sample_type | positive / negative (target present / absent) |
target_speaker | Target speaker id (F1, F2, M1, M2) |
enrollment_lang, mixture_lang | Enrollment vs. mixture language (equal = same-lang, differ = cross-lang) |
snr_db | Target-to-interferer SNR |
enrollment_sentence, target_sentence, interferers | Content and interference metadata |
Speakers
Four speakers, anonymized as F1 and F2 (female) and M1 and M2 (male).
Generation
data_generation/personamix/generate_personamix.py in the Persona-ASR repository regenerates the manifests and mixtures from ASR_88/. The manifests it writes are identical to the released ones, and in our checks the mixtures match the released audio to within one 16-bit sample value. The script's docstring describes the mixing recipe.
License and citation
Released under CC BY 4.0. Please cite Persona-ASR.
@article{meiramov2026personaasr,
author = {Meiramov, Rakhat and Rakhimzhanova, Tomiris and Taibassarov, Adil and Makhataeva, Zhanat and Varol, Huseyin Atakan},
title = {Persona-ASR: Bilingual Target-Speaker Speech Recognition for Kazakh--English Overlapping Speech},
journal = {Machine Learning and Knowledge Extraction},
year = {2026},
volume = {8},
number = {8},
pages = {246},
doi = {10.3390/make8080246}
}
Funding
This research is funded by the Committee of Science of the Ministry of Science and Higher Education of the Republic of Kazakhstan (Grant No. BR24993001).