--- title: "Technical Specifications | Enterprise Speech Data | YPAI" url: https://ypai.ai/speech-data/technical-specifications/ description: "Enterprise speech dataset specifications: audio formats, sample rates, metadata manifests, segmentation, and QA outputs for production ML pipelines." source: "src/copy/routes (route /speech-data/technical-specifications/)" --- # Formats, rates, manifests and QA outputs. The spec sheet your pipeline reads. > Enterprise speech dataset specifications: audio formats, sample rates, metadata manifests, segmentation, and QA outputs for production ML pipelines. WAV or FLAC, 16 to 48 kHz, 16 or 24-bit, structured JSON manifests, per-recording QA status. The project spec sheet is finalised in scoping. - **Format**: WAV · FLAC - **Sample rate**: 16 · 44.1 · 48 kHz - **Bit depth**: 16 · 24-bit - **Channels**: Mono · stereo on request - **Capture**: 48 kHz / 24-bit - **Signal-to-noise**: 40 dB or better - **Segmentation**: Utterance - **Manifest**: JSON, per recording - **Integrity**: SHA-256 per file - **Description**: Croissant 1.1 + data card - **Acceptance**: 98% first-delivery floor - **Pilot**: within 10 business days ## Four numbers the specification carries, measured per recording and reported per delivery. The project sheet sets the rest. Capture floor · acceptance floor · quota tolerance - **Capture sample rate**: 48 kHz. 24-bit, studio and field - **Signal-to-noise**: 40 dB. or better, measured per recording - **Quota tolerance**: 5 pts. per cell, reported each delivery - **First-delivery acceptance**: 98%. floor, re-collected below it ## Constraints are enforced at capture. Invalid audio never enters the pipeline. SNR, noise floor and environment thresholds are set per project. - **Device**: Device class enforced at capture time - **Sample rate**: Matches the project specification before submission - **Noise floor**: Within the agreed threshold, checked automatically - **Clipping**: No clipping or distortion detected - **Duration**: Within the expected range for the prompt - **Rejection**: Non-conforming audio rejected before it enters the pipeline ## Segmented at the utterance, transcribed to your convention, described in the manifest. Recordings are segmented at the utterance level by default. File-level segmentation or another approach is available on request. Transcripts are delivered as verbatim or normalised text, to the project's convention. Label formats follow common training pipeline conventions. One structured JSON record per recording, with the fields in the example record. Full schema documentation is provided in scoping. Additional fields are available on request. manifest.json · one record, example values - **recording_id**: "rec_004217" - **speaker_id**: "spk_0142" - **transcript**: "..." - **duration_ms**: 4210 - **sample_rate**: 48000 - **format**: "wav" - **language**: "nb-NO" - **region**: "Vestlandsk" - **consent_reference**: "cns_8f31" - **qa_status**: "accepted" ## Three QA stages, each one leaving an output in the delivery. Quality threshold enforcement (SNR, clipping, silence detection), synthetic artefact detection, technical specification compliance check, automatic rejection of non-conforming recordings. Review coverage follows the agreed acceptance and sampling plan. The plan may include linguistic correctness, naturalness, script adherence, and per-recording acceptance decisions where applicable. Coverage balance verification, speaker distribution analysis, label integrity check, and final acceptance review before delivery. - **Automated validation**: Per-recording validation result in the manifest. SNR, clipping and silence thresholds, synthetic-artefact detection, specification compliance, automatic rejection of non-conforming recordings. - **Human QA**: Per-recording acceptance decision. Review coverage follows the agreed acceptance and sampling plan, with linguistic correctness, naturalness and script adherence where applicable. - **Dataset-level QA**: QA report with the delivery. Coverage balance, speaker distribution, label integrity, final acceptance review before delivery. ## File-based delivery, versioned, with the handoff procedure in the agreement. Secure transfer, cloud storage handoff or another enterprise-compatible mechanism, agreed per project. Iteration cycles, revision policy and version control are set per project and bound by the contract. Handoff procedures, acceptance criteria, and post-delivery support are documented in the project agreement. ## Delivered for offline training. Folder structure and naming documented per project. YPAI datasets are delivered in formats compatible with standard ML training pipelines. Delivery is file-based and designed for offline training workflows. Datasets delivered in WAV/FLAC with structured JSON manifests, folder structure and naming conventions documented per project, compatible with common ASR/TTS training frameworks. An integration spec is available on request. Delivery package, structure documented per project ## This page describes the conventions. The project specification is written in scoping. - Procurement appendices, legal terms and DPA documentation, linked from the speech data page or provided during scoping ## Formats, metadata and delivery ### What audio formats does YPAI support for speech datasets? YPAI delivers speech datasets in WAV and FLAC formats. The specific format is defined during project scoping based on your pipeline requirements. ### What sample rates are available? Standard sample rates include 16 kHz, 44.1 kHz, and 48 kHz. The appropriate sample rate for your project is determined during technical scoping based on your use case and training requirements. ### How is audio quality validated? YPAI combines automated validation (SNR checks, clipping detection, and technical compliance) with Human QA applied according to the agreed acceptance and sampling plan. The plan defines which recordings receive review before acceptance. ### What metadata is included with delivered datasets? Datasets include structured JSON manifests containing fields such as recording_id, speaker_id, transcript, duration_ms, sample_rate, format, language, region, consent_reference, and qa_status. Additional fields are available on request. ### Can I request stereo recordings instead of mono? Yes. The default configuration is mono, but stereo recordings are available on request and can be specified during project scoping. ### How is data delivered? By secure transfer, cloud storage handoff or another enterprise-compatible mechanism. The chosen method is documented in the project agreement. ### Are YPAI datasets compatible with standard ML frameworks? Yes. Datasets are delivered in formats compatible with standard ML training pipelines, including common ASR and TTS training frameworks. Integration specifications are available on request. ### What QA documentation is provided with delivered datasets? Delivered datasets include QA reports documenting validation results, coverage balance verification, speaker distribution analysis, and label integrity checks. The documentation scope is set per project. ### Can specifications be customised for my project? Yes. Technical specifications are defined per engagement and finalised in scoping. YPAI works with your team to define project-specific requirements, thresholds, and deliverables. ### Is there an API for accessing datasets? YPAI datasets are delivered as file-based packages designed for offline training workflows. Delivery is by secure transfer or cloud storage handoff, agreed in scoping. Bring the pipeline requirement. The spec sheet comes back from scoping. A scoping discussion with the data team sets the project-specific thresholds, formats and deliverables. - [Speech data overview](https://ypai.ai/speech-data/) - [Engagement model](https://ypai.ai/speech-data/engagement-model/) - [Language coverage](https://ypai.ai/speech-data/language-coverage/) - Service Level Agreement - [DPA overview](https://ypai.ai/speech-data/dpa/)