--- license: cc-by-nc-sa-4.0 task_categories: - automatic-speech-recognition language: - sn - en tags: - code-switching - speech - african-languages - asr - maternal-health - shona - zimbabwe size_categories: - n<1K ---

Zambezi ECHO overview

# Zambezi ECHO v1: Shona-English Code-Switched Maternal Health Queries ## Dataset Description Zambezi ECHO v1 SESB (Shona-English Speech Benchmark) is a dataset of short, simulated patient voice queries in Shona (Zimbabwe), code-switched with English, covering common maternal and child health concerns: pregnancy symptoms, danger signs, child illness, and general health questions asked the way patients actually phrase them in the field, mixing Shona with English medical and everyday terms. Unlike a full doctor-patient consultation benchmark, each clip is a single patient utterance or question, not a multi-turn dialogue: the unit here is "how does a patient actually ask this?" rather than a full visit. It spans **25 speakers**, most recorded with the same 31-prompt set, for **779 clips in total** (~3.70 hours of audio; see the duration note under Known Limitations). Speakers vary in accent (Zezuru, Chikaranga, Samanyika), language dominance (Shona-dominant vs. bilingual), age, location, phone type, and recording environment, reflecting realistic field-recording conditions rather than a studio setup. - **License:** CC BY-NC-SA 4.0 - **Clips:** 779 - **Speakers:** 25 (IDs P-001 to P-026; there is no P-023) - **Total duration:** ~3.70 hours (computed over the 750 clips that have a duration value) - **Audio format:** mp3 - **Split:** `train` (single split; see Dataset Structure) ## Dataset Statistics | Speaker | Accent | Profile | Clips | Avg. Duration | Hours | Avg. CMI | Avg. Switch Pts | |---|---|---|---|---|---|---|---| | P-001 | Harare (Zezuru) | Shona-dominant | 31 | 12.4s | 0.107 | 2.8 | 0.5 | | P-002 | Masvingo (Chikaranga) | Shona-dominant | 33 | 17.2s | 0.157 | 35.2 | 9.1 | | P-003 | Harare (Zezuru) | Bilingual | 31 | 24.4s | 0.210 | 11.5 | 5.3 | | P-004 | Masvingo (Chikaranga) | Shona-dominant | 32 | 15.1s | 0.134 | 17.4 | 6.0 | | P-005 | Mutare (Samanyika) | Shona-dominant | 31 | 18.0s | 0.155 | 6.4 | 2.5 | | P-006 | Mutare (Samanyika) | Shona-dominant | 31 | 13.8s | 0.119 | 13.8 | 2.3 | | P-007 | Masvingo (Chikaranga) | Bilingual | 31 | 23.3s | 0.200 | 9.2 | 2.4 | | P-008 | Harare (Zezuru) | Shona-dominant | 31 | 16.3s | 0.140 | 17.2 | 4.3 | | P-009 | Harare (Zezuru) | Shona-dominant | 31 | 19.9s | 0.171 | 23.3 | 5.3 | | P-010 | Masvingo (Chikaranga) | Bilingual | 31 | 26.2s | 0.225 | 31.5 | 11.1 | | P-011 | Harare (Zezuru) | Bilingual | 31 | 18.8s | 0.162 | 5.7 | 2.5 | | P-012 | Harare (Zezuru) | Shona-dominant | 31 | 12.9s | 0.111 | 5.7 | 1.3 | | P-013 | Masvingo (Chikaranga) | Shona-dominant | 31 | 14.3s | 0.123 | 7.9 | 1.8 | | P-014 | Harare (Zezuru) | Shona-dominant | 31 | 17.3s | 0.149 | 4.8 | 1.8 | | P-015 | Mutare (Samanyika) | Bilingual | 32 | 21.5s | 0.191 | 10.0 | 3.6 | | P-016 | Mixed | Shona-dominant | 31 | 15.5s | 0.133 | 10.7 | 2.4 | | P-017 | Mutare (Samanyika) | Shona-dominant | 31 | 14.3s | 0.123 | 5.5 | 1.1 | | P-018 | Masvingo (Chikaranga) | Shona-dominant | 31 | 19.2s | 0.165 | 7.9 | 3.2 | | P-019 | Harare (Zezuru) | Bilingual | 31 | 17.6s | 0.151 | 6.5 | 2.1 | | P-020 | Mutare (Samanyika) | Bilingual | 31 | 14.1s | 0.121 | 29.3 | 5.9 | | P-021 | Harare (Zezuru) | Bilingual | 31 | 24.5s | 0.211 | 22.3 | 7.9 | | P-022 | Harare (Zezuru) | Shona-dominant | 31 | 20.0s† | 0.011 | 10.8 | 2.1 | | P-024 | Harare (Zezuru) | Bilingual | 31 | 16.4s | 0.141 | 10.9 | 3.1 | | P-025 | Gweru | Shona-dominant | 31 | 14.3s | 0.123 | 6.9 | 2.0 | | P-026 | Harare (Zezuru) | English--dominant | 31 | 18.9s | 0.163 | 21.8 | 5.8 | | **Total** | | | **779** | **17.8s** | **3.70** | **13.4** | **3.8** | P-022 has a duration value for only 2 of its 31 clips, so its average duration and hours are not representative. The Total row's duration and hours are computed over the 750 clips that have a duration. **CMI** (Code-Mixing Index) and **switch points** are computed from the `[[EN]]`/`[[MED]]` tag boundaries in the transcription (see Notes; this is a heuristic, not the validated Das & Gambäck toolkit). The Total row shows the clip-level mean across all 779 clips. `[SHR]` spans do not appear to be counted as English in these values (see Notes). ### Code-switching - **589 of 779 clips (76%)** contain at least one English span; **190 clips (24%)** have no `[[EN]]` tag. - **241 clips (31%)** contain at least one `[[MED]]` (medical terminology) span. - **33 clips (4%)** contain at least one `[SHR]` span (41 spans in total) marking Shona words that are direct derivatives of English words, for example *screeniwa* (from "screening"). These come from 7 speakers (P-015, P-016, P-017, P-018, P-019, P-025, P-026). Of these, 6 clips contain `[SHR]` spans but no `[[EN]]` span. - 184 clips (24%) have no English-origin tag of any kind (`[[EN]]` or `[SHR]`). - Clip-level CMI ranges from 0 to 50 (median 8.9); the mean is 13.4, and the mean number of switch points per clip is 3.8. - The amount of code-switching varies substantially by speaker, from near-zero (P-001, avg. CMI 2.8) to heavy mixing (P-002, avg. CMI 35.2). ### Speaker composition | Attribute | Breakdown (speakers) | |---|---| | Accent | Harare (Zezuru): 12 · Masvingo (Chikaranga): 6 · Mutare (Samanyika): 5 · Mixed: 1 · `Gweru`: 1 [speaker recorded Gweru as an accent when they meant Ndebele] | | Language profile | Shona-dominant: 15 · Bilingual: 9 · `English--dominant` [sic]: 1 | | Age | 19-47 (mean 29.0, median 27) | | Recording environment | Home: 20 · School: 2 · Indoors: 1 · Outdoors: 1 · | | Session noise level | Moderate: 14 · Quiet: 11 | ### Clip durations Clips with a duration value (750 of 779) average 17.8 s (median 16.4 s, min 7.7 s, max 101.0 s) and contain on average 5.5 annotated speech segments across all 779 clips. Transcriptions average about 27 words per clip. ## Dataset Structure Each example is one clip (metadata is provided in `metadata.csv`): - `file_name`: relative path to the audio file (e.g. `data/P-011_Q4.mp3`) - `audio`: the recorded utterance (mp3), HuggingFace `Audio` feature - `patient_id`: anonymized speaker identifier (P-001 … P-026, no P-023) - `clip_id`: prompt identifier within that speaker's session (Q1-Q31 for most speakers; P-002 has Q1-Q33, and P-004 and P-015 have Q1-Q32) - `transcription`: verbatim human transcription (plain text) - `transcription_tagged`: same transcription with English spans wrapped in `[[EN]]`…`[[/EN]]`, medical terminology spans additionally wrapped in `[[MED]]`…`[[/MED]]` (matching the AfriSwitchCare tagging convention), and Shona words that are direct derivatives of English words wrapped in `[SHR]`…`[/SHR]` (see Tagging Scheme below) - `duration`: clip duration in seconds (blank for 29 clips of P-022) - `num_speech_segments`: number of annotated speech regions in the clip - `cmi`: approximate Code-Mixing Index (see Notes) - `num_switch_points`: approximate count of language alternations - `age`: speaker age at recording time - `location`: speaker's recorded location (town/suburb) - `accent`: speaker's regional Shona accent (`Harare (Zezuru)`, `Masvingo (Chikaranga)`, `Mutare (Samanyika)`, or `Mixed`; P-025 is recorded as `Gweru`) - `language_profile`: `Bilingual` or `Shona-dominant` (P-026 is recorded as `English--dominant`) - `phone_type`: device used to record (varies across speakers) - `recording_environment`: e.g. `Home`, `School`, `Indoors`, `Outdoors` - `noise_level`: `Quiet` / `Moderate` (session-level) - `health_literacy`: speaker's self-reported health/medical background (free text, e.g. nutritionist, data analyst, O-level biology) - `consent_given`: recording consent flag (`True` for all 779 clips) - `timestamp`: recording timestamp (sessions recorded 12-14 August 2026) There is a single `train` split; most speakers recorded the same 31 prompts, so downstream users can carve out their own dev/test split by speaker or by prompt as needed. ### Tagging scheme | Tag | Meaning | Example | |---|---|---| | `[[EN]]`…`[[/EN]]` | English span | `kuita [[EN]]weak[[/EN]] nekuva` | | `[[MED]]`…`[[/MED]]` | Medical terminology (nested inside `[[EN]]` when the term is English) | `[[EN]][[MED]]pharmacy[[/MED]][[/EN]]` | | `[SHR]`…`[/SHR]` | Shona word that is a direct derivative of an English word | `[SHR]screeniwa[/SHR]` (from "screening") | `[SHR]` marks Shona-ised English (English stems with Shona morphology), as opposed to `[[EN]]`, which marks English words and phrases used as English. ## Dataset Creation ### Source Data Simulated patient voice queries recorded in the field across several Zimbabwean towns (Harare including Ruwa and Domboshava, Masvingo, Gweru, Mutare, Rusape, Bulawayo areas) and annotated in Label Studio with speech/noise region labels and verbatim transcriptions. English, medical-term and Shona-derivative (`[SHR]`) spans were tagged inline by annotators as they transcribed. This release contains transcriptions and audio only. ## Notes / Known Limitations - **CMI and switch points are approximate.** They are computed heuristically from the `[[EN]]`/`[[MED]]` tag boundaries (word-count ratio and tag-transition count between the two), not from a validated code-switching toolkit like the one used for AfriSwitchCare (Das & Gambäck 2014 / Gambäck & Das 2016). Treat these fields as indicative rather than exact. - **`[SHR]` is not reflected in CMI or switch points.** The `cmi` and `num_switch_points` values do not count `[SHR]` spans as English. They will understate mixing for the 33 clips that contain `[SHR]` (most heavily for P-026, P-016 and P-018). - **Tag formatting is not fully consistent.** `[SHR]` uses single brackets (unlike `[[EN]]`/`[[MED]]`), many tags have spaces inside or around the spans (e.g. `[SHR] screeniwa [/SHR]`), and a few clips contain malformed tags (e.g. `[ENor]`, `[ENso]`, `[ENconfusion]`, `[/END]`, extra closing brackets such as `[[/EN]]]`). Normalize and validate tags before parsing. - **Accent and location can differ.** `accent` is the speaker's self-reported regional accent and `location` is where they were recorded or live, so they do not always match (e.g. P-011 has a Harare (Zezuru) accent but was recorded in Masvingo; P-026 has a Harare (Zezuru) accent but was recorded in Gweru; P-020 has a Mutare (Samanyika) accent but was recorded in Harare). P-016 is labelled `Mixed`, and P-025's accent is recorded as `Gweru` (a location, recorded in Harare). - **Free-text metadata fields are not normalized.** `location`, `phone_type`, `recording_environment` and `health_literacy` contain inconsistent spellings and spacing as entered (e.g. `Quet` for `Quiet`, `O lvel`, `Olevel`, `Sumsang A5`, trailing spaces). Normalize before grouping on these fields. `language_profile` for P-026 is entered as `English--dominant`. - **Missing durations.** 29 clips (P-022 Q2-Q30) have a blank `duration`, so duration statistics and hours cover 750 of the 779 clips. - **Duration outlier.** One clip (P-010 Q31) is 101 s, far above the 17.8 s mean (the next longest is 41.8 s). ## Licensing Information Released under **Attribution-NonCommercial-ShareAlike 4.0 (CC BY-NC-SA 4.0)**.