Skip to content
A Top 5 global AI outsourcing company by Outsource Accelerator, above Scale AI.
Speech AI · East Africa · Central Asia · Speech & audio datasets

Swahili & Uzbek TTS cut word error rate from 11.2% to 4.3%

A speech team's models underperformed in Swahili and Uzbek because available data was English-first and dialect-blind.

11.2%
WER before
4.3%
WER after
2 dialect-deep
Languages
East Africa · Central Asia
Region

The challenge and pain points

Off-the-shelf datasets missed regional pronunciation and code-switching, driving a high word error rate in production.

  • Word error rate in Swahili and Uzbek was high enough that voice features were not shippable in those markets.
  • Off-the-shelf datasets were English-first and missed regional pronunciation, so models trained on them failed on real speech.
  • Natural code-switching between local languages and English was absent from available data, so the model stumbled on mixed-language utterances.
  • Dialect variation was flattened into a single generic voice, which did not match how people in each region actually speak.
  • Buying more third-party data did not help, because the gap was coverage, not volume.

What Corpshore did

Recorded and transcribed studio-grade TTS/ASR data with native speakers in Kampala and Tashkent, covering regional dialects and natural code-switching.

  1. 1
    Define the dialect and code-switch spec

    Set the target dialects and the mixed-language patterns to cover in Swahili and Uzbek before any recording began.

  2. 2
    Recruit native speakers in-region

    Staffed native speakers in Kampala and Tashkent who live in the languages, rather than diaspora or second-language voices.

  3. 3
    Capture studio-grade TTS and ASR audio

    Recorded controlled, studio-grade audio for TTS and baseline ASR, with prompts written to elicit regional pronunciation and natural code-switching.

  4. 4
    Transcribe and label under the QA cascade

    Native transcribers produced aligned transcripts with the three-tier QA cascade gating each batch for accuracy before delivery.

  5. 5
    Validate against the target languages

    Measured word error rate on the target languages so the improvement was verified on the exact markets the model needed to serve.

The solution

Corpshore built the dataset from capture through transcription rather than assembling it from third-party sources. Native speakers in Kampala and Tashkent recorded studio-grade audio to a spec that deliberately included regional pronunciation and the code-switching people use in everyday speech.

Every recording was transcribed and aligned by native speakers, then gated through the three-tier QA cascade so the training set was clean from the source. Dialect coverage was defined per language so the data matched the markets the team serves, not a generic approximation.

Trained on this purpose-built data, the model's word error rate on the target languages fell from 11.2% to 4.3%, which moved voice features from not shippable to production-ready in those markets.

Results

Word error rate on the target languages fell from 11.2% to 4.3%, a reduction of roughly 62%, which was the difference between voice features that could not ship and ones that could. The gain came from covering the regional pronunciation and code-switching that English-first datasets miss, captured by speakers who live in the languages. Because the data was collected in-region rather than bought at a premium, the program also carried Corpshore's typical cost advantage versus US-domestic collection.

MetricResultNotes
Word error rate, before11.2%On target languages, off-the-shelf data
Word error rate, after4.3%On target languages, purpose-built data
Relative reduction≈62%Computed from the before and after figures
Languages2 dialect-deepSwahili and Uzbek
SpeakersNative, in-regionKampala and Tashkent
CoverageDialect + code-switchMissing from English-first data
QAThree-tier cascadeGated each batch before delivery
Word error rate on target languages
11.2%
Before
4.3%
After
62%
reduction

Real delivered figures. Word error rate on Swahili and Uzbek fell from 11.2% to 4.3% after training on the purpose-built dataset.

Where errors get caught (three-tier QA cascade)
Tier 1 · annotator + peer~80%
Tier 2 · expert QA lead~15%
Tier 3 · programmatic + consensusremainder

Representative distribution of caught errors across the cascade, locking in 97%+ accuracy before delivery.

Client names and some figures are confidential. Where an exact client metric is not published, outcomes are described qualitatively and charts show representative or methodology-level data, including the three-tier QA cascade distribution and Corpshore's typical 50 to 70% cost advantage versus US-domestic providers.

Ready to scope a pilot?

Tell us your modality, volume, and languages. We'll return an indicative scope, timeline, and cost band.

Start a pilot Explore careers