The value model
One recording. Six datasets.
A single consented story take self-labels into six kinds of ground truth a lab would otherwise assemble from six separate vendors. Nothing extra is annotated by hand — the structure falls out of how the take is captured.
1
Transcript / ASR
The read text IS the reference transcript — supervised speech-to-text, aligned to audio.
2
Forced alignment
Per-sentence segments give time-stamped units — training data for aligners and duration models.
3
Accent label
The speaker's L1 and country stamp the recording — a labeled accent example for robustness.
4
Code-switch boundary
In Bantu&English takes, the switch point (switch_ms) is a labeled language boundary.
5
TTS voice
Clean, consented, page-length native speech — training material for a natural voice.
6
Prosody / phrasing
Connected page-length reading carries phrasing and rhythm that isolated units cannot.
Per mode
Each mode decomposes differently.
The same six-way idea, specialised by mode — this is what a machine sees for a real take in each mode (LIVE = extractable today, PENDING = on the roadmap).
Bantu only
bem · Ulubuli Lulaletelela · 3.5s
asr
audio span ms: [0, 3540] · segmented: True
tts voice
audio span ms: [0, 3540] · sr hz: 48000 · accent: bem/Zambia
aligned units
n segments: 0 · method: read-aloud
translation pair
present: True
prosody
LIVE
connected: True
lexicon fsi
PENDING
Bantu & English (code-switch)
bem · Imbali Shibili · 6.4s · switch @ 4588 ms
asr
audio span ms: [0, 6557] · segmented: True
tts voice
audio span ms: [0, 6557] · sr hz: 48000 · accent: bem/Zambia
aligned units
n segments: 2 · method: read-aloud
translation pair
present: True
codeswitch boundary
LIVE
switch ms: 4588
prosody accent
LIVE
connected: True
segment preview (first 2)
English only (Bantu-accented)
lug · The Morning at Home · 253.4s
asr
audio span ms: [0, 253440] · segmented: True
tts voice
audio span ms: [0, 253440] · sr hz: 48000 · accent: lug/Uganda
aligned units
n segments: 92 · method: read-aloud/stt-aligned
accent label
LIVE
accent: lug/Uganda
wer benchmark
LIVE
wer: 0.0802
prosody
PENDING
connected: True
segment preview (first 8)
Oral narration
loz · None · 174.2s
asr
audio span ms: [0, 174170] · segmented: True
tts voice
audio span ms: [0, 174170] · sr hz: 48000 · accent: loz/Zambia
aligned units
n segments: 0 · method: read-aloud
translation pair
present: False
prosody
LIVE
connected: True
lexicon fsi
PENDING
Reproducibility
The same text, across accents.
Mode 3 uses a small set of language-neutral English stories read by many Bantu L1 speakers. Same words, different mouths — a matched benchmark that isolates accent from content. That is why the accent WER numbers are comparable across languages.