My agent is pronouncing em-dashes in responses saying, "A-T" or "O-T" or "OT"

Support Help
1

Hi Retell support team,

My customer service agent is using and pronouncing out loud em-dashes saying, “A-T” or “O-T” or “OT”

In the after call summary I can see the use of em-dashes where the agent says one of the above. An example was when I was testing and asked my agent where to find documentation on setting up an AI assistant. Her response:

Great — here are the quickest places to get setup guidance: - Getting Started with Athena AI Assistant (step‑by‑step for campaigns, contacts, workflows) — available in the Resources & guides section of the platform.

It sounded like, “Great A-T, here are the quickest places to get setup guidance…”

Its driving me mad and nothing I have tried works. I have:

  • Added specific prompt language under my Communication Style and Behavior heading, “Never utilize or say em-dashes aloud”
  • In the Speech Settings, under Pronunciation, I added an entry for em-dash but left the pronunciation blank.

Do you have ideas on how to resolve this?

Thank you in advance,

Jeff

2

The issue you’re experiencing is likely related to how the LLM is generating text with em-dashes, which the TTS system then attempts to pronounce. Here are several approaches to resolve this:

1. Enable Speech Normalization

Turn on the normalize_for_speech setting, which converts text to its spoken form before audio generation(1)(2)(3). This helps the TTS system handle punctuation and special characters more naturally(1)(2)(3).

You can enable this in your agent settings or via the API(2)(3).

2. Add Prompt Instructions

Include specific guidance in your agent’s prompt about avoiding em-dashes in responses(4). Try adding something like:

When responding, never use em-dashes (—) in your text. Instead, use commas, periods, or start new sentences to separate thoughts.

(4)

3. Configure Pronunciation Dictionary

While you’ve tried adding an entry with blank pronunciation, you may need to explicitly define how the em-dash should be handled(1)(5)(2)(3). The pronunciation dictionary accepts entries in this format(2)(3):

{
  "word": "—",
  "alphabet": "ipa",
  "phoneme": ""
}

(2)(3)

However, a better approach might be preventing the LLM from generating em-dashes in the first place through prompt engineering(4).

4. Adjust Voice Settings

Consider testing different voice models, as some handle punctuation better than others(1)(5)(2)(3). You can select different voice models in your agent’s voice settings(1)(5)(2)(3).

The most effective solution will likely be combining speech normalization with clear prompt instructions to avoid em-dashes altogether(1)(4).

3

Hello @digitalbar

Can you share the call id with the timestamp?

5

Hi Omar,

Agent ID: agent_c3cf02cde168b2cc1f2059aa78

Call ID: call_1c000423a2410c7b31e780c11dc

6

Oh - the Timestamp - 0:43

7

Thank you for providing that additional context about the timestamp. This helps clarify that the em-dash pronunciation issue is occurring at a specific point in your call (0:43 seconds in).

Based on the available documentation, here’s my refined recommendation:

Primary Solution: Speech Normalization

Enable the normalize_for_speech setting(1)(2)(3). This setting specifically converts punctuation and special characters to their spoken form before audio generation, which should prevent the TTS from attempting to pronounce the em-dash character(1)(2)(3).

Secondary Solution: Prompt Engineering

Add explicit instructions in your agent’s Global Prompt to avoid using em-dashes entirely(1)(2). For example:

Never use em-dashes (—) or en-dashes (–) in your responses. Use commas, periods, or semicolons instead to separate thoughts.

Why Your Current Attempts May Not Be Working

Your pronunciation dictionary entry with blank pronunciation may not be effective because the TTS system still needs guidance on what to do with the character(1)(2)(3). The normalization setting is specifically designed to handle these text-to-speech conversion issues before the audio generation phase(1)(2)(3).

If the issue persists after enabling speech normalization and updating your prompts, you may want to reach out through the official support channels for additional troubleshooting specific to your agent configuration.

8

Hello @digitalbar

I see you have a pronunciation configured for the em-dashes. Can you remove it and try again? Also can you try a different voice provider?

10

I only added that pronunciation to try and correct the issue. It is not the reason for it. The AI suggested prompt update has mitigated the issue a bit. I will also try a different voice provider.

11

Try a different voice provider, That should solve it

12

Switching voice providers can help address audio detection issues in Retell AI(1). When experiencing problems with audio not being picked up, you can configure a different voice provider for your agent(2)(1).

To change the voice provider, update the voice_id parameter when configuring your agent(2). For example, you could switch from ElevenLabs to OpenAI voices(2)(1).

Additionally, you can set up TTS fallback by configuring the fallback_voice_ids parameter(2)(3). The fallback voice IDs must be from different TTS providers than your primary voice(2)(3). The system will go through the fallback list in order if the primary provider experiences issues(2)(3).

For audio recognition problems specifically with Italian or other languages, the team has indicated this may be related to ASR (Automatic Speech Recognition) issues with short sentences(1). Using boosted keywords can help improve specific word transcription(2)(1). You can add keywords to the boosted_keywords parameter, which biases the transcriber model to more likely transcribe those words(2)(3).