I tried the Solaria model from Gladia because it was supposed to be more accurate and have a smaller latency than deepgram. I did some testing and ended up having an average of 2.42s when using Solaria compared to 1.24s for deepgram nova-2-general (exact same parameters for the rest) I don't understand the gap?? https://cdn.discordapp.com/attachments/1395407673016717375/1395407673343606826/Screenshot_2025-07-17_at_16.11.30.png?ex=687a562a&is=687904aa&hm=21b6c984ff78d13fb6d888fb68fa9673222514676d30dec3e68f4e9c7c9e4b47&

Looking into it

could you also share: When exactly this happened (the timestamp) What response you expected to get What response you actually got instead This would really help us figure out what went wrong! @pierreh

FYI, this is what they said: Hi, thanks for reaching out and for testing Gladia! You're absolutely right to expect lower latency and improved French accuracy with Solaria, that’s what we designed it for. That said, I don’t know the exact implementation details of how Gladia is integrated within Vapi, and from what I could find in their documentation, it’s unclear how they pass parameters like endpointing or custom vocabulary. This could be part of the latency discrepancy you're observing. Here are a few suggestions and examples you can try to improve your setup: ✅ Recommended Gladia Parameters Please make sure you're explicitly setting these parameters in your config (either directly or by asking Vapi support if these are passed through properly): Endpointing Lowering this value can drastically reduce perceived latency: { "endpointing": 0.05 } 📘 Docs on endpointing Custom Vocabulary If your use case includes domain-specific terms, brand names, or unusual phrases, you can use custom vocabulary to improve accuracy: "realtime_processing": { "custom_vocabulary": false, "custom_vocabulary_config": { "vocabulary": [ "Westeros", { "value": "Stark" } ], "default_intensity": 0.5 } 📘 Docs on custom vocab Speech Threshold Tuning (for noisy or degraded calls) If your calls are 8kHz or have background noise, try: { "pre_processing": { "audio_enhancer": false, "speech_threshold": 0.6 } } 📘 Docs on speech threshold

Hey pierreh, I sincerely apologize for the delay in our responses this week. We experimented with some process adjustments that didn’t work as intended, and unfortunately, this caused some disruptions. That’s entirely my responsibility. Just a quick note—our team is unavailable on weekends. However, if something urgent or critical comes up, feel free to reach out to me directly, and I’ll assist you. The support issues will be resolved in the coming weeks, and starting Monday, you can expect more consistent and improved support. Thank you so much for your patience and for continuing to stick with us!

Thank you @Shubham Bajaj , have a great WE!

We reached out to our team to have them take a look at our current Gladiia implementation and see what they can do to improve the experience for our users. Thank you for your patience while our team reviews it

thanks!

Great ! Adding on this the requested feature of custom vocabulary : https://vapi.canny.io/feature-requests/p/add-custom-vocabulary-feature-to-gladia-transcriber, also essential to proper gladia integration

Our team is currently working on making improvements to our implementation for Gladia. While there is no exact ETA on its release, you will either hear from us when it's deployed or it'll show up in the changelogs for upcoming updates to the platform.

any news on that ?

@pierreh I've recently asked during an OH (vapi-stage) and @Shubham Bajaj told me it was planned for last week / this week I believe 🤔 I hope to get some news soon !

No update yet, but it is in the works. Thank you for your patience

It seems they are even going one step further in the nearby future, for even lower latency ! https://cdn.discordapp.com/attachments/1395407673016717375/1405113525923483810/image.png?ex=689da572&is=689c53f2&hm=3ff5bdc73da59f35a83c8ad0db45d7ecc3ac81930b9b9221b1d4a60400af6045&

It should be live if it was found in our August 3rd changelog. Feel free to check it out

I've done a few tests but all transcriptions are translated in English by Gladia by default (couldn't change that). Opened a new ticket for this : https://discord.com/channels/1211482211119796234/1405511793064939562

let me know @Jeebs from eva.be when you here from them! We are getting there

@pierreh it does work 👌🏼

what is your first feedbacks on Gladia vs Deepgram @Jeebs from eva.be ? Do you see a difference?