DOCSEARCHbyAlgolia
Submit Docs

Speechmatics Documentation

website·Indexed 37 days ago

https://docs.speechmatics.com

Documentation for Speechmatics speech-to-text services, including the ASR REST API (v2.0.0), Realtime API, and Management API (v1.0.0). Covers cloud and on-premise deployments via Kubernetes containers and Virtual Appliances, GPU inference, language identification, and workspace administration including SSO, billing, and usage reporting.

Tokens
91.6K
Snippets
447
Records
640
Agent score
69%

What's inside Speechmatics

  1. Overview of the Batch persistent worker

    The Batch persistent worker (or HTTP batch worker), available from version 15.7.0, is a long-running transcription service. Unlike standard batch containers that exit after one job, the persistent worker loads ASR models once at startup and accepts multiple jobs via an HTTP API. This eliminates per-job cold starts and allows for concurrent processing using the --parallel flag to manage processing units (engines).

  2. Overview of Agent STT for voice agents

    Agent STT is a SaaS-only transcription service designed for voice agents. It provides turn-based, speaker-labelled transcription that is punctuated and ready for LLM consumption. Key features include:

    • Turn detection: Identifying when a speaker has finished talking.
    • Intelligent segmentation: Grouping words into speaker-attributed segments.
    • Diarization: Labeling different speakers.
    • Speech signals: Notifications for the start and end of speech.

  3. Overview of the Voice Agent API

    The Voice Agent API is a preview WebSocket API designed for building voice agents. It provides speaker-labelled, turn-based transcription that is punctuated and ready for LLM consumption. Turn detection is handled server-side based on the selected profile.

    Important Constraints:

    • This is a preview offering and should NOT be used for live production traffic.
    • There are no uptime, performance SLAs, or data residency guarantees (processing may occur in US and EU regions).

  4. Organize work using Projects

    Projects are organizational units within a workspace used to scope API keys and track usage. API keys, transcripts, and voice agent configurations belong to a specific project. Any member of the workspace can access any project within it.

    Key constraints:

    • Each workspace has one default project that cannot be deleted.
    • A workspace can have up to 1000 projects.
    • API keys are scoped to a single project; a key from one project cannot access transcripts from another.
    • Billing is not segmented by project, but usage can be viewed per project.

  6. Compare Speechmatics Cloud and On-prem deployment options

    Speechmatics offers two primary deployment methods:

    1. Cloud (SaaS): A fully managed, scalable service. Users can get started via the Speechmatics Portal and access Realtime transcription, Batch transcription, and Agent STT.
    2. On-prem: Deployment via containers in a private environment for maximum control over data and infrastructure. Available containers include CPU and GPU speech-to-text engines, Kubernetes services, Language ID containers, and Translation containers.

  7. Use Melia 1 Realtime Preview for multilingual transcription

    Melia 1 is an early preview model for the Realtime API that provides code-switching capabilities for 56 languages. It automatically detects the spoken language and labels each word with the detected language.

    Key Constraints:

    • This is for evaluation only and is not intended for production use.
    • Only a subset of Realtime features is currently supported.
    • Full production release is expected in October 2026.

    Access:

  9. Compare Standard Batch vs Persistent Worker

    The Persistent Worker is optimized for high-throughput workloads and short audio files by eliminating the 'cold start' overhead of loading ASR models (which can take several seconds). While standard batch containers launch per job, the persistent worker is a single long-lived container where multiple jobs share resources.

    FeatureStandard BatchPersistent Worker
    Startup costPer jobOnce
    Memory usageOne container per jobMultiple jobs share one container
    CPU/GPU utilizationInterrupted between jobsContinuous
    Best forLarge, infrequent filesHigh throughput or smaller files

  10. Detect end of speech for turn detection

    The 'End of Utterance' feature allows developers to detect when a person has finished speaking in real-time. This is primarily used for Voice AI, real-time translation, and dictation to trigger system responses or finalize transcriptions without awkward delays.

  11. Understand the Speechmatics JSON transcript response schema

    The default JSON transcript response (version 2.1) contains three primary top-level objects:

    1. job: Summary of the ASR job including id, created_at, data_name, and duration.
    2. metadata: Details about the output generation, including type (usually transcription) and transcription_config.
    3. transcription_config: The configuration used for the job, including:
      • language: ISO language code.
      • model: The model used (standard, enhanced, or melia-1).
      • diarization: Speaker/channel labeling (none, speaker, or channel).
      • punctuation_overrides: Settings for sensitivity (0 to 1) and permitted_marks.
      • additional_vocab: Custom words/phrases for recognition.
      • enable_entities: Boolean to include entities like dates and numbers.

  12. Detect topics in English audio transcriptions

    Speechmatics allows the detection of topics within audio files using a Large Language Model (LLM). This feature is currently only supported for English. Users can provide a custom list of up to 10 topics to detect. Because it uses an LLM, results may occasionally be inaccurate or vary across multiple runs of the same file.