Recognition settings

Set language on every request, because Rime does not detect the language. Rime transcribes in the language that you select, so if you send Spanish speech as en, you get a poor transcript, not an error. Two more settings are optional: the mode sets the form of the text, and context terms help the model recognize names.

Setting	Required	Values	Default
Language	Yes	`en` for English or `es` for Spanish	No default. Set it on every request.
Mode	No	Written or word-for-word	Written
Context terms	No	Names and phrases that you expect to hear	No terms

The three settings work the same way in the SDK, the WebSocket API, and both gRPC methods.

Word-for-word mode

The mode tells the model what form of text to write.

	Written mode (default)	Word-for-word mode
Numbers	As digits. Codes are in compact form.	As spoken words
Fillers, repetitions, and false starts	The mode does not ask the model to keep them or to remove them.	The mode asks the model for each one.
Choose it for	Text that you show, or send to a language model	An exact record of the words. Examples are calls that you review for quality or compliance, and speech analytics.

The quickstart transcribed the same synthesized call in both modes:

Mode	Final transcript
Written (default)	`Hi! I'm calling about flight 247 to Denver. Can I, move it to Friday?`
Word-for-word	`Uh hi. I'm calling about flight two four seven to Denver, can I um move it to Friday?`

Word-for-word mode kept “Uh”, “um”, and the spoken “two four seven”. Written mode wrote 247, and on this recording it also left out both fillers. Even so, the mode guides the model but does not guarantee the form of each word. Test the mode on recordings like the ones that your application will get. Phone audio is one such case. When callers read out ZIP codes, phone numbers, or account numbers, use word-for-word mode. In Rime’s tests on phone audio, it kept every digit. Set word-for-word mode in the stream call:

Python
Node.js
Go
Rust

from rimelabs_sdk import TranscriptionMode

transcript = client.stt.stream(audio, language="en", mode=TranscriptionMode.VERBATIM)

const transcript = client.stt.stream(audio, { language: "en", mode: "verbatim" });

transcript, err := client.STT.Stream(ctx, source, rime.TranscriptionOptions{
    Language: "en",
    Mode:     rime.TranscriptionVerbatim,
})

use rimelabs_sdk::{TranscriptionMode, TranscriptionOptions};

let transcript = client
    .stt()
    .stream(source, TranscriptionOptions::new("en").mode(TranscriptionMode::Verbatim))?;

Wire values for direct API calls

Set mode to TRANSCRIPTION_MODE_VERBATIM in the WebSocket start message, or in the gRPC config or options.TRANSCRIPTION_MODE_WRITTEN is the default. If you omit mode or send TRANSCRIPTION_MODE_UNSPECIFIED, the service uses written mode. An unknown value fails with invalid_input over WebSocket or INVALID_ARGUMENT over gRPC.

Language

Use en for English or es for Spanish, and choose the language from what you already know about the speaker. For example, use the caller’s menu choice or account settings.

Language errors and confirmation

The service checks the tag before it accepts audio. Each of these values fails with invalid_input over WebSocket or INVALID_ARGUMENT over gRPC:

You send	Why it fails
No `language`, or `und`	A language is required
A supported tag in the wrong form, such as `en_US`	It is not a valid BCP 47 tag
Another language, such as `fr`	Only `en` and `es` are supported

The SDKs send language without changes, and the service checks it.The direct API returns the confirmed language in started.accepted.language over WebSocket and in accepted.language over gRPC. The final result also has it in language. Its source is always LANGUAGE_SOURCE_SELECTED.

Context terms

Context terms tell the model which names and phrases to expect in this utterance. For example, an agent for flight changes can send the caller’s surname from the booking, such as Okonkwo. The agent can also send its own airline and fare names, such as Northwind Air and FlexPlus. Terms are hints without weights, so they make a match more likely, but they do not force a match. Add the terms to the stream call:

Python
Node.js
Go
Rust

transcript = client.stt.stream(
    audio,
    language="en",
    context_terms=["Northwind Air", "FlexPlus", "Okonkwo"],
)

const transcript = client.stt.stream(audio, {
  language: "en",
  contextTerms: ["Northwind Air", "FlexPlus", "Okonkwo"],
});

transcript, err := client.STT.Stream(ctx, source, rime.TranscriptionOptions{
    Language:     "en",
    ContextTerms: []string{"Northwind Air", "FlexPlus", "Okonkwo"},
})

let terms = vec!["Northwind Air".into(), "FlexPlus".into(), "Okonkwo".into()];
let transcript = client
    .stt()
    .stream(source, TranscriptionOptions::new("en").context_terms(terms))?;

Test terms on your own recordings, and keep the terms that change the result. Terms share the model’s context with the audio, so a long list leaves less space for a long utterance.

Context term rules

If you omit the list or send an empty list, the service adds no hints. A term that is empty or has only whitespace fails with invalid_input.
The service keeps the order and the text of each term, including spaces and commas in the term. It joins the terms with a comma and a space.
The joined list, with its separators, must fit in the limit of 1,024 UTF-8 bytes, not characters. There is no limit on the number of terms.
A term cannot contain the control markers of the model, such as text between <| and |>. A term with a control marker fails with invalid_input.
The service rejects input that is too large with invalid_input. It never truncates the input.

Field names in the direct APIs

Direct API calls use these fields for the three settings:

API	Language	Mode	Context terms
WebSocket `start`	`language`	`mode: "TRANSCRIPTION_MODE_VERBATIM"`	`contextTerms`
gRPC `config` or `options`	`language`	`mode: TRANSCRIPTION_MODE_VERBATIM`	`context_terms`

Each SDK documents its options and their defaults in the rime-sdk repository: Python, Node.js, Go, and Rust.

Get started

Build with Rime

Integrations

Account & updates

Recognition settings

Word-for-word mode

Language

Context terms

Field names in the direct APIs

​Word-for-word mode

​Language

​Context terms

​Field names in the direct APIs

Word-for-word mode

Language

Context terms

Field names in the direct APIs