For this call: 7e1ac148-751a-4988-8acf-a4d7db664b65 There is a language pronunciation issue. Call is in french, and the whole conversation is in french. But Agent pronounces 2025 in english. When checking the logs for this call, we can see that voice input sent to Cartesia is:

Voice input: Voici quelques créneaux disponibles pour un rendez-vous :. - Jeudi 9 janvier two thousand twenty five :

Instead of correct french:

Voice input: Voici quelques créneaux disponibles pour un rendez-vous :. - Jeudi 9 janvier deux mille vingt cinq :

Model output contains

202 5

, how is it converted to

two thousand twenty five

? Logs:

17:04:46:706
[LOG]
Model request started (attempt #1, gpt-4o-2024-08-06-global, azure-openai, westus)

17:04:47:269
[LOG]
Model output: quelques

17:04:47:269
[LOG]
Model output: Voici

17:04:47:277
[LOG]
Model output: cré

17:04:47:286
[LOG]
Model output: ne

17:04:47:294
[LOG]
Model output: aux

17:04:47:304
[LOG]
Model output: disponibles

17:04:47:312
[LOG]
Model output: pour

17:04:47:322
[LOG]
Model output: un

17:04:47:339
[LOG]
Model output: -vous

17:04:47:339
[LOG]
Model output: rendez

17:04:47:472
[LOG]
Model output: Je

17:04:47:472
[LOG]
Model output: -

17:04:47:472
[LOG]
Model output: :

17:04:47:473
[LOG]
Model output: 9

17:04:47:473
[LOG]
Model output: janvier

17:04:47:473
[LOG]
Model output: udi

17:04:47:473
[LOG]
Model output:

17:04:47:473
[LOG]
Model output:

17:04:47:474
[LOG]
Model output: 202

17:04:47:477
[LOG]
Model output: 5

17:04:47:477
[LOG]
Model output: :

17:04:47:478
[LOG]
Model output:

17:04:47:479
[LOG]
Voice input: Voici quelques créneaux disponibles pour un rendez-vous :. - Jeudi 9 janvier two thousand twenty five :

@marem It seems there was a misinterpretation during the transcription process. The model output "202 5" should have been transcribed as "deux mille vingt cinq" to correctly reflect the French pronunciation of the year 2025. The transcription error resulted in the English pronunciation "two thousand twenty five" being used instead. To resolve this issue, we need to adjust the transcriber settings to ensure that numbers are transcribed in the correct language context. Since the transcriber is currently set to 'deepgram', the model to 'Nova 2', and the language to English, we should verify if there's an option to handle multilingual content or specifically transcribe numbers in the language of the conversation, which in this case is French. If there's a setting or parameter that can be adjusted to correct this behavior, it should be implemented. Otherwise, this might be a limitation of the current transcriber configuration, and further support from the transcriber service provider might be necessary. For now, please let me know if you would like to proceed with adjusting the transcriber settings or if you need further assistance with this issue.

We use deepgram, nova 2 and fr language

@marem Thank you for clarifying the use of Deepgram, Nova 2, and the French language setting. It appears there may be an issue with the language model or the transcription service not correctly interpreting the numerical sequence "202 5" as "deux mille vingt cinq." To address this, we should review the transcription settings to ensure that numbers are being transcribed in the correct language context. If the settings are correct, we may need to report this issue to the transcription service provider for further investigation. In the meantime, if there are any immediate actions or adjustments you would like to make, please let me know how I can assist you further.

logs

🔵 16:04:47:479 Voice Input Formatted: "Voici quelques créneaux disponibles pour un rendez-vous :. - Jeudi 9 janvier two thousand twenty five :", Original: "Voici quelques créneaux disponibles pour un rendez-vous : - Jeudi 9 janvier 2025 :" @marem This issue is related to voice input formatting. Currently, there is no option to enable or disable this feature, but I will consult with the team to explore possible solutions. The system automatically formats numbers and converts them into English words. As a workaround, you can modify your prompt to generate numbers in French instead of numeric format. https://cdn.discordapp.com/attachments/1326317518843023482/1326505992095141948/Screenshot_2025-01-08_at_16.29.30.png?ex=677fac79&is=677e5af9&hm=17d5effef1123d4fd1d4432093c904e75258d07a8888ae60a127c5137f309c61&

ok, good to know. Which are the rules of the voice input formatting process? For instance, why 9 is not converted to nine in "Jeudi 9 janvier two thousand twenty five"? That can help me to know which number I need to generate in french

@marem By default, the system uses the voice format plan, which is a component of the voice chunk plan. Since no specific format was provided, it follows the default rules where numbers less than 10 are not formatted. For more information about the format plan, please refer to the API documentation under the voice chunk plan section. https://cdn.discordapp.com/attachments/1326317518843023482/1326514148770713631/Screenshot_2025-01-08_at_17.02.56.png?ex=677fb412&is=677e6292&hm=85823bc20d607a8b0ca5f4bfc855934a9acf4a6d15867a55acf52532b44c243c&

ok, I can see that in the assistant api reference: https://docs.vapi.ai/api-reference/assistants/create

"chunkPlan": {
      "enabled": true,
      "minCharacters": 30,
      "punctuationBoundaries": [
        "。",
        "，",
        ".",
        "!",
        "?",
        ";",
        "،",
        "۔",
        "।",
        "॥",
        "|",
        "||",
        ",",
        ":"
      ],
      "formatPlan": {
        "enabled": true,
        "numberToDigitsCutoff": 2025
      }
    },

Is there a documentation somwehere to explain what chunckPlan is for and what we can do with it? is it a new feature? When I check the assistant payload used for the call, there is no mention of chunkplan, so I guess as you said I've no control on it yet?

@marem will add documentation for it, and make it available in the dashboard. If this issue is resolved for you can i close this ticket?