tarirozw/Zambezi_ECHO_v1

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 15 speakers, most recorded with the same 31-prompt set, for 468 clips in total (~2.3 hours of audio). 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: 468
• Speakers: 15
• Total duration: ~2.29 hours
• Split: train (single split; see Notes)

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-020	Mutare (Samanyika)	Bilingual	31	14.1s	0.121	29.3	5.9
Total			468		2.29	14.9 (avg)

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). 133 of 468 clips (28%) contain no English code-switching at all (Shona only); 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).

Dataset Structure

Each example is one clip:

• audio: the recorded utterance (mp3), HuggingFace Audio feature
• patient_id: anonymized speaker identifier (P-001 … P-020)
• clip_id: prompt identifier within that speaker's session (most speakers Q1–Q31; a couple of speakers recorded a small number of extra prompts)
• transcription: verbatim human transcription (plain text)
• transcription_tagged: same transcription with English spans wrapped in [[EN]]…[[/EN]], and medical terminology spans additionally wrapped in [[MED]]…[[/MED]] (matching the AfriSwitchCare tagging convention)
• duration: clip duration in seconds
• 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
• language_profile: Bilingual or Shona-dominant
• phone_type: device used to record (varies across speakers)
• recording_environment: e.g. Home
• noise_level: Quiet / Moderate (session-level, may be blank where not recorded)
• health_literacy: speaker's self-reported health/medical background (e.g. nutritionist, data analyst, O-level biology)
• consent_given: recording consent flag
• timestamp: recording timestamp

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.

Dataset Creation

Source Data

Simulated patient voice queries recorded in the field across several Zimbabwean towns (Harare, Masvingo, Gweru, Mutare, Rusape, Bulawayo areas) and annotated in Label Studio with speech/noise region labels and verbatim transcriptions. English and medical-term 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.

Licensing Information

Released under Attribution-NonCommercial-ShareAlike 4.0 (CC BY-NC-SA 4.0).