Izwe.ai

Converts audio input into text across all 11 official South African languages (Zulu, Xhosa, Sotho, Tswana, Venda, Tsonga, Afrikaans, English, Ndebele, Swati, and Sepedi) using language-specific acoustic models and phonetic training data optimized for regional dialects and pronunciation patterns. The platform likely employs language detection to automatically identify the spoken language or allows manual language selection, then routes audio through language-specific ASR (automatic speech recognition) pipelines rather than using generic multilingual models.

Accepts audio and video file uploads through a web interface or API endpoint, queues them for asynchronous transcription processing, and returns completed transcripts via webhook callbacks or polling. The system likely implements a job queue (Redis, RabbitMQ, or similar) to manage concurrent transcription requests, with worker processes handling the actual ASR computation. Upload handling probably includes file validation, format detection, and optional compression for bandwidth optimization.

Analyzes audio input to automatically identify which of the 11 supported South African languages is being spoken, then routes the audio to the appropriate language-specific ASR model without requiring manual language selection. This likely uses a lightweight language identification (LID) classifier running on audio spectrograms or MFCC features, with fallback to manual language selection if confidence is below a threshold. The routing mechanism ensures that Zulu speech doesn't get processed by an English model, preserving accuracy.

Indexes completed transcripts for full-text search, allowing users to query across transcription archives by keyword, phrase, or language. The platform likely builds inverted indices (Elasticsearch, Solr, or similar) for each language, with language-specific tokenization and stemming rules to handle morphological complexity in Bantu languages. Search results probably return matching transcript segments with timestamps, enabling users to jump directly to relevant audio sections.

Exports completed transcripts in multiple formats (plain text, SRT/VTT subtitles, JSON, CSV, DOCX) with optional formatting options like timestamp inclusion, speaker labels, and language metadata. The export pipeline likely includes format-specific serialization logic, with subtitle formats (SRT/VTT) handling timestamp synchronization and character limits per line. JSON export probably includes structured metadata (language, confidence scores, speaker info) for downstream processing.

Provides REST API endpoints for developers to integrate transcription capabilities directly into custom applications, with authentication via API keys, request/response in JSON format, and support for both synchronous polling and asynchronous webhook callbacks. The API likely follows RESTful conventions (POST /transcribe, GET /jobs/{id}, etc.) and may include rate limiting, request signing, and detailed error responses. Developers can submit audio URLs or file uploads, specify language preferences, and retrieve results programmatically.

Provides per-word or per-segment confidence scores indicating the ASR model's certainty in the transcription output, allowing users to identify potentially inaccurate sections. The system likely computes confidence as a probability score (0-1) from the acoustic model's output probabilities, with aggregation to segment or sentence level. High-confidence sections (>0.95) are likely accurate, while low-confidence sections (<0.70) may require manual review or re-processing with different settings.

Attempts to identify and label different speakers in multi-speaker audio, segmenting the transcript by speaker with labels like 'Speaker 1', 'Speaker 2', or ideally speaker names if provided. Diarization likely uses speaker embedding models (x-vectors, speaker verification networks) to cluster similar voices and assign consistent labels across the transcript. This is particularly useful for interviews, meetings, and panel discussions where multiple voices are present.

Ensures transcribed audio and text data remain within South African jurisdiction for regulatory compliance, likely storing data in local data centers and implementing audit logging for access and processing. The platform probably handles POPIA (Protection of Personal Information Act) compliance requirements, including data retention policies, deletion on request, and consent management. Audit trails track who accessed transcripts and when, supporting compliance verification and incident investigation.

Offers pricing in South African Rand (ZAR) with payment methods common in South Africa (EFT, credit cards, potentially mobile money), and billing structures tailored to local business needs. The platform likely avoids the premium pricing of global competitors by operating locally, reducing currency conversion costs and payment processing fees. Billing may support monthly or usage-based models with transparent per-minute or per-hour transcription rates.