Speechmatics

Converts live audio streams to text with claimed sub-1-second latency using a proprietary neural acoustic model optimized for streaming inference. Supports continuous audio input via persistent connections (WebSocket or gRPC streaming), with intermediate results returned before final transcription is complete, enabling responsive voice interfaces and live captioning without perceptible delay.

Processes pre-recorded audio files (WAV, MP3, Opus, etc.) asynchronously, returning full transcriptions with optional domain-specific vocabulary via custom dictionary. Supports up to 10 concurrent file jobs per second (Pro tier), with job queuing and async completion callbacks (webhook mechanism unconfirmed). Custom dictionaries allow injection of domain terminology (e.g., medical terms, product names) to reduce transcription errors in specialized contexts.

Secures API access via API key authentication (format unspecified; likely 'Authorization: Bearer' or 'X-API-Key' header). Enforces tier-based rate limits and monthly quotas: Free tier (480 min/month STT, 1M chars/month TTS, 2 concurrent sessions), Pro tier (480 min/month free + overage, 50 concurrent sessions, 10 file jobs/sec), Enterprise (unlimited). Rate limits prevent abuse and ensure fair resource allocation across users.

Freemium pricing model offering 480 minutes/month of speech-to-text transcription and 1M characters/month (~20 hours) of text-to-speech synthesis without credit card requirement. Enables developers to prototype and test Speechmatics APIs before committing to paid tiers. Free tier includes 2 concurrent real-time sessions and English-only TTS. Overage usage requires upgrade to Pro or Enterprise tier.

Startup incentive program offering up to $50k in API credits for early-stage companies, reducing cost of speech recognition and synthesis during product development and scaling. Application-based program (criteria and approval timeline not documented). Credits likely apply to all API usage (STT, TTS, custom models) and may have expiration dates or usage restrictions.

Provides a paid tier at $0.24 per hour of transcription with a 20% discount available for volume commitments. The Pro tier includes 480 minutes of free monthly transcription (matching free tier) plus overage billing, 50 concurrent sessions for real-time transcription, and 10 file jobs per second for batch processing. Pricing structure and overage rates are not fully documented.

Recognizes speech in 55+ languages and language variants using a single unified multilingual acoustic model, with optional automatic language detection (no pre-specified language code required) or explicit language specification. Supports code-switching (mixing languages within a single utterance) and regional variants (e.g., British English, Mandarin vs. Cantonese). Language detection likely uses a classifier on initial audio frames to route to appropriate language-specific decoder.

Specialized acoustic and language model trained on medical terminology, clinical dictation, and healthcare-specific speech patterns. Reduces transcription errors on medical terms by up to 50% (claimed) compared to general-purpose model through domain-specific vocabulary, acoustic adaptation, and likely medical-specific language model decoding. Intended for clinical documentation, medical transcription services, and healthcare voice applications.

Automatically detects speaker boundaries and identifies distinct speakers in multi-speaker audio (e.g., conversations, meetings, interviews) without requiring pre-defined speaker profiles. Uses speaker embedding models (likely x-vector or speaker-encoder based) to cluster speech segments by speaker identity, outputting transcription with speaker labels (e.g., 'Speaker 1:', 'Speaker 2:'). Supports 2-N speakers with no documented upper limit.

Converts text to natural-sounding speech with claimed low latency suitable for real-time voice agent interactions. Supports English language (with 'more languages coming soon'). Synthesis likely uses a neural vocoder (e.g., WaveGlow, Glow-TTS) for naturalness and fast inference. Optimized for voice agent use cases where response latency directly impacts perceived responsiveness (target: <500ms for typical agent responses).

Enterprise-tier capability enabling development of custom synthetic voices tailored to organization branding, speaker identity, or specific use cases. Likely involves voice cloning or speaker adaptation techniques (e.g., speaker embedding fine-tuning, speaker-conditional TTS) using organization-provided audio samples. Custom voices can be deployed on-premises or in private cloud for regulated environments. Implementation details (training data requirements, adaptation time, voice quality metrics) not documented.

Enables deployment of Speechmatics speech recognition and synthesis models on customer-managed infrastructure (on-premises data centers, private cloud, edge devices) for organizations with data residency, compliance, or latency requirements. Supports air-gapped deployments with no external API calls. Likely includes containerized model packages (Docker), licensing mechanisms, and optional hardware acceleration (GPU support). Eliminates cloud dependency and enables compliance with HIPAA, GDPR, and other data protection regulations.

Provides precise word-level timestamps and audio alignment, enabling synchronization of transcription with video, subtitles, or other time-based media. Returns start/end timestamps for each word (and optionally phoneme-level timing) with confidence scores. Useful for video captioning, subtitle generation, and audio-visual synchronization. Enterprise-tier feature with higher accuracy and finer granularity than standard transcription.

Translates transcribed speech or text to target languages, enabling cross-lingual communication and content localization. Likely uses neural machine translation (NMT) models trained on multilingual corpora. Can be applied post-transcription (transcribe in source language, then translate) or as part of unified transcription-translation pipeline. Supports 55+ language pairs with varying translation quality depending on language pair and domain.