Lexia's Gilbert API provides a comprehensive RESTful interface for integrating speech recognition capabilities into enterprise applications. The API design emphasizes developer ergonomics while maintaining the flexibility needed for diverse use cases. Our API handles everything from simple synchronous transcription requests to complex streaming scenarios with real-time partial results.
Authentication and authorization form the foundation of secure API access. Gilbert API supports multiple authentication methods: API keys for simple integrations, OAuth 2.0 with PKCE for web applications, and mTLS (mutual TLS) for server-to-server communication requiring the highest security. We implement rate limiting per API key to prevent abuse, with configurable quotas that scale with subscription tiers. Enterprise customers can configure custom rate limits aligned with their expected usage patterns.
The core transcription endpoint accepts audio in multiple formats: PCM, FLAC, MP3, WAV, and WebM. Audio is automatically transcoded to the optimal format for our models (16kHz mono PCM for most use cases). We support chunked uploads for large files, enabling streaming uploads that begin processing before the entire file is received. This reduces end-to-end latency significantly for long recordings.
Request parameters allow fine-grained control over transcription behavior. The `language` parameter enables explicit language specification, important for multilingual scenarios. Our models support over 50 languages with varying accuracy levels—English and French achieve WER < 5%, while less-common languages may see WER around 8-10%. The `model` parameter selects between base models optimized for different use cases: `general` for conversational speech, `medical` for healthcare terminology, `legal` for legal proceedings, and `technical` for engineering and scientific content.
Advanced features include speaker diarization, which separates multiple speakers in a single audio stream. Our diarization system uses spectral clustering on speaker embeddings extracted from the audio, achieving speaker change detection accuracy above 95% for conversations with distinct speakers. This is particularly valuable for meeting transcription where identifying 'who said what' is essential.
Timestamps and word-level confidence scores provide additional metadata for downstream processing. Each word in the transcription includes start and end timestamps (millisecond precision), enabling synchronization with video or creating searchable, time-indexed transcripts. Confidence scores range from 0-1, allowing applications to flag low-confidence segments for human review. We've found that segments with confidence < 0.7 benefit significantly from manual correction.
Streaming API endpoints enable real-time transcription for live applications. Clients establish WebSocket connections and send audio chunks continuously. The API returns partial transcriptions as speech is processed, updating previous segments as context improves. This enables live captioning, real-time meeting notes, or interactive voice interfaces. Our streaming implementation maintains sub-200ms latency for partial results, crucial for conversational applications.
Error handling and retries are essential for production API usage. Gilbert API returns standard HTTP status codes and detailed error messages. Transient failures (network issues, temporary service unavailability) should trigger exponential backoff retries. We recommend implementing idempotency keys for critical requests to prevent duplicate processing if retries are necessary. Our API guarantees exactly-once processing when idempotency keys are used.
Cost optimization strategies include intelligent caching, batch processing when real-time results aren't required, and selective use of premium features. Our pricing model charges per audio minute processed, with volume discounts for enterprise customers. Features like speaker diarization and custom model inference incur additional costs, so applications should use these selectively based on value. We provide usage analytics dashboards that help identify optimization opportunities.