I've made a Voice Agent that acts like a receptionist in german. I am actually pretty satisfied with the results but there is 1 problem that needs to get fixed or get better: The reliability from transcribers in general. I need to make sure, that the transcriber gets the output from the user as accurate as possible. I am using deepgram as the provider and the nova 2 model. So far, this got me the best results, but it still get's some words wrong, especially Names and E-Mails. Even tough the user have to spell their names & emails. My question is: what are the best options, to finetune given transcriber to the highest degree of "reliability". I know 99% is unrealistic, but maybe 90% would be awesome. Best regards

Hey @allain222 To help track down this issue, could you share: - The call ID - When exactly this happened (the timestamp) - What response you expected to get - What response you actually got instead This would really help me figure out what went wrong!

Okay: Call ID: 16680ad5-1450-41de-8c59-ce5ed3359519 1st Timestamp: on Minute 00:48 the Agent asked the user for a birthdate. User told the Agent: 26.03.1940; the uttered words from the user were "sechsundzwanzigster dritter neunzehnhundertvierzig" The Transcriber wrote "neunzehnhundertneunzehn" instead of "neunzehnhundertvierzig" The Agent however told the user, that he noted her birthdate as "26.03.1999" instead of "26.03.1940." 2nd Timestamp: 01:00 Agent ask the user for her email-adress. The user replies with "goldmeier@web.de" The transcriber wrote: goldmeier Web und e. After the reply from the user, the Agent ask (correctly) the user to spell out her email. The user replies with: "g-o-l-d short pause "m-a-i-e-r @ web.de" The transcriber wrote: "N a I e at BETT Punkt. Siehe, er." The Agent therefore repeats the transcribers text, which confuses the user because it did not resemble the email in any shape of form. Thanks for the fast reply, would really appreciate the help. Everything else works perfect, it really is just this transcriber issue which effects the reliability from the product. Best regards Elias

@allain222 Upon reviewing your call ID, I have identified two potential solutions for email-issue involving prompting. I will outline both approaches below 1. Request that the user spell their email address one character at a time. For instance, instruct them to articulate: "g - o - l - d - m - e - i - e - r, dot, d - e." After they have spelled it, kindly repeat it back for confirmation: "I heard 'g-o-l-d.' Is that correct? Please confirm or correct me." 2. Encourage users to adhere to a structured input process. For email addresses, prompt them to articulate each component clearly and separately, using "at" for "@" and "dot" for "." For example: "Please state your email address letter by letter. For '@,' say 'at,' and for '.', say 'dot.'"

Sorry for the delayed response. First of all, thanks for the help and quick responses. After reading your 2 potential solutions, a few questions came into my mind: Approach Nr. 1: For approach number one, the prompt structure you suggested me I already have implemented. I instruct the agent to always ask the user to spell out their E-mail AND Name. As you could hear on the call, the Agent 1st asked the user for their email, and IF they could spell it out. The user responded by saying out their email without spelling it, the agent acknowledged, that the user did not spell out her email as he kindly asked, therefore asking the user again, to spell out the email. The user then spelled her email letter by letter in an understandable manner: "g-o-l-d short pause "m-a-i-e-r @ web.de" The transcriber understood this spelling as "N a I e at BETT Punkt. Siehe, er." Please let me know if I undersstood your approach correctly. -------------------------------------------------------------------- Approach Nr.2: So to understand correctly, the user from given example spelled out her email too fast? That's why I should tell the Agent, to tell the user to articulate each letter clearly? If that's the case, I definitely will definitely change the agent's prompt.

> So to understand correctly, the user from given example spelled out her email too fast? No that's not the case but can be used as an suggestion for the user, I suggest trying out different providers as well because transcriber do has issues with capturing emails correctly. > Approach Nr. 1: I will watch for alternatives to capture email correctly because it has to be a single solution for almost all of the users.