Buyer problem
Speech datasets become unreliable when scripts are misread, audio is clipped or noisy, speaker metadata is incomplete, consent is unclear, or recognition errors are measured without Hebrew-specific linguistic analysis.
Suitable use cases
- Scripted sentence recording
- Prompt-based or domain-specific speech collection
- ASR evaluation and regression testing
- Transcription QA
- Speaker and recording-condition coverage analysis
Inputs required from the client
- Script or prompt set
- Speaker profile and coverage requirements
- Consent and permitted-use language
- Audio format, sample rate, and naming rules
- Acceptance and rejection criteria
Method and workflow
Speaker instructions, consent records, recording requirements, and rejection rules are established. A pilot validates script clarity and audio quality before recruitment or production expands.
Deliverables
- Audio files and transcripts
- Speaker and recording metadata
- Consent status records
- Audio-quality and completeness checks
- Rejected-item reasons and QA summary
- ASR error analysis where requested
Acceptance criteria and quality metrics
Metrics may include script accuracy, signal quality, clipping/noise rate, metadata completeness, transcription accuracy, word error rate, critical-term error rate, and speaker-distribution coverage.
Example output
The public sample batch acceptance summary demonstrates how batch status and rejection reasons can be reported.
Data-handling considerations
Voice data requires explicit consent and usage scope. The project must define whether recordings are used for training, evaluation, research, or commercial deployment, along with retention and transfer rules.
Request a Hebrew speech or ASR pilot
Start with a small speaker and script batch to validate instructions, audio settings, consent tracking, metadata, and rejection rules.