Gladia

Processes pre-recorded audio files through an asynchronous queue-based system that routes requests across multiple AI transcription engines (including the proprietary Solaria model) to optimize for accuracy across 100+ languages. The system handles variable audio durations, supports concurrent processing up to tier-specific limits (25 concurrent for Starter, unlimited for Enterprise), and returns time-stamped transcripts via REST API with optional webhook callbacks for completion notification.

Provides WebSocket-based live transcription of audio streams with claimed sub-300ms latency, enabling real-time caption generation and voice AI agent interactions. Supports concurrent streaming connections (30 for Starter, unlimited for Enterprise) with automatic language detection and code-switching across multiple languages within a single stream. Integrates natively with voice infrastructure platforms (LiveKit, Pipecat, Vapi) via pre-built connectors.

Automatically segments long audio recordings into chapters or topics based on content analysis, generating chapter markers with timestamps and titles. Enables navigation of long-form content (podcasts, lectures, interviews) by breaking them into logical sections. Implementation approach (automatic vs. manual, algorithm used) not documented.

Provides native integration with SIP (Session Initiation Protocol) telephony systems and legacy phone infrastructure, with audio optimization for 8 kHz sample rate (standard for telephony). Enables real-time transcription of phone calls without requiring intermediate recording or forwarding services. Supports both inbound and outbound call transcription with automatic call metadata capture (caller ID, duration, etc.).

Provides native connectors and SDKs for popular voice AI frameworks (Pipecat, Vapi, LiveKit) and no-code automation platforms (Zapier, Make, n8n), enabling one-line integration without raw API implementation. Pre-built connectors handle authentication, connection pooling, error handling, and reconnection logic. Supports both async and real-time transcription modes through framework-specific abstractions.

Implements a usage-based pricing model where customers pay per hour of audio processed (not per request or per token), with tiered pricing based on monthly commitment level (Starter: $0.61/hr async, $0.75/hr real-time; Growth: $0.20/hr async, $0.25/hr real-time with 67% discount; Enterprise: custom). Concurrency limits scale by tier (25 async/30 real-time for Starter, unlimited for Enterprise). Starter tier includes 10 free hours/month.

Offers contractual zero-data-retention guarantees for Enterprise tier customers, ensuring audio files and transcripts are not stored, used for model training, or retained after processing. Provides full data sovereignty with compliance certifications (GDPR, HIPAA, AICPA SOC 2 Type II claimed). Growth+ tiers offer automatic model training opt-out; Enterprise has default opt-out. Enables deployment in regulated industries without data residency concerns.

Automatically segments audio into speaker turns and labels each segment with a speaker identifier (Speaker 1, Speaker 2, etc.), enabling multi-speaker conversation analysis. Works across both async and real-time transcription modes, identifying speaker boundaries through audio analysis without requiring pre-registered speaker models or enrollment. Output includes speaker labels in transcript timestamps and optional speaker confidence scores.

Detects the language(s) spoken in audio automatically without requiring pre-specification, supporting transcription in 100+ languages with code-switching capability (handling mid-sentence language switches). Uses the Solaria model and multi-engine routing to optimize accuracy across linguistic families (Indo-European, Sino-Tibetan, Afro-Asiatic, etc.). Returns detected language codes and per-segment language labels in transcript output.

Allows injection of domain-specific terminology, proper nouns, and custom spellings into the transcription model to improve accuracy for specialized vocabularies (medical, legal, technical, brand names). Accepts a vocabulary list at request time and applies it during transcription to boost recognition of custom terms. Implementation details (vocabulary size limits, matching algorithm, priority over base model) not documented.

Automatically extracts named entities (person names, email addresses, phone numbers, organizations, locations) and personally identifiable information (PII) from transcribed audio. Operates as a post-processing step on generated transcripts, identifying and optionally redacting sensitive data. Supports entity classification (PERSON, ORG, LOCATION, EMAIL, PHONE, etc.) with confidence scores.

Generates summaries of transcribed audio content, condensing long conversations or recordings into concise text summaries. Supports both abstractive summarization (generating new summary text) and extractive summarization (selecting key sentences from transcript). Integrates with LLM backends for abstractive mode, with optional custom summarization prompts. Works on both async and real-time transcription outputs.

Generates time-stamped subtitle files (SRT, VTT formats) from transcribed audio, enabling video captioning and accessibility. Automatically segments transcript into subtitle chunks with appropriate timing based on speaker pacing and natural sentence boundaries. Supports customizable subtitle duration and character limits per line.

Provides direct integration between transcribed audio and large language models (LLMs) without requiring intermediate API calls or third-party service orchestration. Routes transcripts directly to LLM backends (OpenAI, Anthropic, or proprietary) for downstream processing (summarization, entity extraction, sentiment analysis, etc.). Supports custom prompts and model selection per request.

Analyzes the emotional tone and sentiment of transcribed audio, classifying speech into sentiment categories (positive, negative, neutral, mixed) with confidence scores. Operates as a post-processing step on transcripts, analyzing speaker tone, word choice, and linguistic patterns. Supports per-speaker sentiment analysis for multi-speaker conversations.