For AI agents: a documentation index is available at /llms.txt. Markdown versions of all pages can be requested by appending `.md` to the URL, or by setting the `Accept` header to `text/markdown`.
Skip to main content
Speech to Text

Models

Speechmatics offers four models for Speech to Text. All four use the same API. The model you choose determines accuracy, how the model handles more than one language, and which features you can configure.

Choose a model​

Your interaction pattern determines which models you can use, so start there.

Models for pre-recorded transcription​

Standard, Enhanced, and Melia 1, all generally available.

  • Enhanced — the highest accuracy on single-language audio, such as medical, legal, or subtitling work. The only model offering the medical domain.
  • Standard — when throughput or latency matter more than maximum accuracy, such as archival transcription, content indexing, or large-scale captioning.
  • Melia 1 — audio that contains more than one language, including speakers who switch language mid-conversation.

To compare them feature by feature, see pre-recorded transcription features.

Models for streaming transcription​

Standard, Enhanced, and Melia 1. Standard and Enhanced trade accuracy against throughput exactly as they do for pre-recorded audio, and both are generally available.

Melia 1 for streaming is in Preview, on a separate endpoint. See Melia 1 for streaming.

To compare them feature by feature, see streaming transcription features.

Models for agent STT​

Linden 1 only, in Preview. Use agent STT for conversational applications that need a turn-based transcript ready to pass to a language model.

For what it supports, see agent STT features.

Specify a model​

Set the model property in your transcription config. If you do not set it, the standard model is used.

This config selects the enhanced model:

{
  "type": "transcription",
  "transcription_config": {
    "model": "enhanced",
    "language": "en"
  }
}

Linden 1 is the exception: you select it by connecting to an agent STT endpoint rather than by setting model.

Standard​

Standard prioritizes throughput. It transcribes a single language, selected with a transcription language pack, and it can detect the language automatically.

Enhanced​

Enhanced delivers the highest accuracy of the four models. Like Standard, it transcribes a single selected language and can detect the language automatically.

Enhanced also offers the medical domain, which tunes recognition for healthcare audio. Standard does not.

Melia 1​

Melia 1 is a multilingual model. It transcribes audio that contains more than one language, including speakers who switch language mid-conversation, and returns a single continuous transcript. It does not require you to select a language pack, and its accuracy is on par with Standard.

Set "model": "melia-1" and "language": "multi":

{
  "type": "transcription",
  "transcription_config": {
    "model": "melia-1",
    "language": "multi"
  }
}

Melia 1 does not support the auto language value, which returns an error. Set language to multi.

Melia 1 has no language pack selection. For pre-recorded transcription on SaaS on Cloud it is served from the EU1 and US1 endpoints:

RegionEndpoint
EU1 (Europe)eu1.asr.api.speechmatics.com
US1 (USA)us1.asr.api.speechmatics.com

For the full list of endpoints, see Authentication. For the regions and features available with Melia 1, see Feature availability.

Melia 1 for streaming​

Streaming with Melia 1 is available on SaaS on Cloud for evaluation and feedback. It is not production-ready and not ready to scale.

Streaming with Melia 1 runs on a dedicated Preview endpoint, wss://preview.rt.speechmatics.com/v2, and returns a language label on every word. On this endpoint Melia 1 transcribes multilingual audio more accurately than Enhanced, and monolingual audio more accurately than Standard.

For the configuration, the supported settings and a runnable example, see Mixed-language transcription.

Linden 1​

Linden 1 serves agent STT only. It returns transcripts as complete speaker turns rather than as a word-by-word stream, with turn detection running server-side, which suits applications that hand each turn to a language model.

Linden 1 returns segment-level timings and segment-level partials rather than the word-level equivalents the other three models return.

Linden 1 is available on SaaS on Cloud for evaluation and feedback. It is not production-ready and not ready to scale.

Deprecated: operating_point​

The model property replaces the operating_point property. Existing configs that use operating_point continue to transcribe without changes.

On SaaS on Cloud, operating_point is deprecated. It maps to model and accepts the same enhanced and standard values. Use model going forward.