The right parameter configuration can make a significant difference in transcription quality and latency for realtime use cases. This guide covers recommended starting points for common scenarios and highlights pitfalls that frequently trip up new integrations.
These recommendations apply to the Realtime
API and are passed during session
initialization. They are starting points — tune them to match your
specific needs.
Language Configuration
One of the most common configuration mistakes is misunderstanding how language_config works. Choosing the right setup avoids unnecessary detection overhead and improves accuracy.
When to set an explicit language:
- You know the language of the audio ahead of time.
- The audio is monolingual (single language throughout).
- You want the fastest, most accurate results.
{
"language_config": {
"languages": ["en"],
"code_switching": false
}
}
- You process audio in many different languages and don’t know which one beforehand.
- You want Gladia to pick the language automatically.
{
"language_config": {
"languages": [],
"code_switching": false
}
}
When code_switching is false and no language is set, the language is
detected on the first utterance and reused for the rest of the session or
file. If the beginning of your audio contains silence, music, or a different
language than the main content, this can lead to incorrect detection for the
whole transcription.
Even when using auto-detection, pass a small list of likely languages in
languages to constrain the search. This improves both accuracy and
processing time.
Code Switching
Code switching (language_config.code_switching: true) lets Gladia detect and transcribe multiple languages within the same audio, re-evaluating the language on each utterance.
When to enable it:
- Speakers switch languages mid-conversation (e.g. bilingual meetings, multilingual customer support).
- You need the detected
language returned per utterance.
When NOT to enable it:
- The audio is in a single language — code switching adds unnecessary processing and can introduce misdetections.
- You’ve set exactly one language in
languages — in that case code_switching is ignored anyway.
{
"language_config": {
"languages": ["en", "fr", "es"],
"code_switching": true
}
}
Do not enable code_switching with an empty languages list. When no
languages are specified, the language detector evaluates every utterance
against 100+ supported languages, which leads to frequent misdetections —
especially between similar-sounding languages. Always provide a short list of
languages you actually expect in the audio.
Custom Vocabulary
Custom vocabulary is a post-transcription replacement based on phoneme similarity. It’s essential for domain-specific terms that speech models frequently mis-transcribe.
Best practices:
- Always provide both the
custom_vocabulary flag and a custom_vocabulary_config.
- Add pronunciations for words that can be said in different ways (accents, foreign speakers). This is more reliable than raising
intensity.
- Keep
intensity moderate (0.4-0.6). High values increase false positives where unrelated words get replaced.
- Set
language on individual vocabulary entries when your audio is multilingual and a term is pronounced differently depending on the language.
For simple terms that are already close to their phonetic spelling (e.g. brand
names), you can pass them as plain strings instead of objects — Gladia will
use the default intensity.
{
"audio_url": "YOUR_AUDIO_URL",
"custom_vocabulary": true,
"custom_vocabulary_config": {
"vocabulary": [
"Kubernetes",
{
"value": "Gladia",
"pronunciations": ["Gladya", "Gladiah"],
"intensity": 0.5
},
{
"value": "PostgreSQL",
"pronunciations": ["Postgres Q L", "Post gress"],
"intensity": 0.4
}
],
"default_intensity": 0.5
}
}
Voice Agents
For callbots, customer-service assistants, or voice-driven chatbots the top priority is low latency. The agent must react quickly to user speech, even if sentence boundaries are not perfectly formed.
| Parameter | Recommended value | Why |
|---|
endpointing | 0.05 - 0.1 | Closes utterances fast, keeping turn-taking snappy. See Endpointing. |
maximum_duration_without_endpointing | 15 | Prevents very long utterances from staying open without cutting off the conversation. |
messages_config.receive_partial_transcripts | true | Enables interim results so the agent can start processing early. Use the speech_stop event to know when the user has finished speaking. See Partial transcripts. |
realtime_processing.custom_vocabulary | true | Add product names and action keywords so the agent can react accurately. |
This setup is optimized for fast turn-taking. If utterances get cut off
mid-sentence, raise endpointing slightly.
Meeting Recorders
For apps that record and transcribe meetings in real time — team stand-ups, board sessions, 1-on-1s — the goal is to produce a structured, speaker-attributed live transcript that can feed downstream features like summarization or live note-taking.
| Parameter | Recommended value | Why |
|---|
endpointing | 0.3 - 0.5 | Lets speakers finish their sentences before closing an utterance. See Endpointing. |
maximum_duration_without_endpointing | 15 | Prevents very long utterances in case a speaker doesn’t pause. |
messages_config.receive_partial_transcripts | true | Feeds live captions to the UI while waiting for final results. See Partial transcripts. |
language_config.languages | Set explicitly | Meeting language is almost always known in advance — setting it avoids detection overhead. |
realtime_processing.custom_vocabulary | true | Add company-specific terms, project names, and participant names for better accuracy. |
Diarization vs. multi-channel: if each speaker is on a separate audio channel (e.g. a, use the channel field on each utterance to identify who is speaking — diarization is not needed. See Multiple channelsIf all speakers share a single audio channel, enable diarization to separate the speakers. See Speaker diarization.
Call Centers
For live phone calls the priorities are speaker identification and fast, accurate transcription despite variable audio quality (telephony codecs, background noise, cross-talk).
| Parameter | Recommended value | Why |
|---|
endpointing | 0.2 - 0.4 | Keeps turn-taking responsive without cutting off mid-sentence. See Endpointing. |
maximum_duration_without_endpointing | 15 | Prevents very long utterances in monologue-style segments. |
language_config.languages | Set explicitly (e.g. ["en"]) | Call center audio typically has a known language. Setting it avoids detection errors on noisy recordings. |
realtime_processing.custom_vocabulary | true | Add product names, plan names, and internal terminology. |
Diarization vs. multi-channel: if each speaker is on a separate audio channel (e.g. a, use the channel field on each utterance to identify who is speaking — diarization is not needed. See Multiple channelsIf all speakers share a single audio channel, enable diarization to separate the speakers. See Speaker diarization. For calls with more than two participants (e.g. conference bridges), use
diarization_config.min_speakers / max_speakers instead of number_of_speakers to give the
model a flexible range.
Subtitles & Captioning
When providing live subtitles, the goal is to sync text with the speaker in real time. The right balance between speed and segment quality depends on whether captions are displayed live or post-produced.
| Parameter | Recommended value | Why |
|---|
endpointing | 0.3 (live) / 0.8 (post-production) | Lower values keep captions close to the speaker; higher values produce cleaner subtitle blocks. |
maximum_duration_without_endpointing | 5 | Prevents excessively long subtitle segments that are hard to read on screen. |
messages_config.receive_partial_transcripts | true | Shows words as they are spoken, then refines them when the final result arrives. |
language_config.languages | Set explicitly | Avoids detection lag when the broadcast language is known. |
For post-production subtitles generated from a recording, consider using the
Pre-recorded API with the dedicated
subtitles feature instead — it
produces SRT/VTT files with fine-grained timing controls. 
