How do I implement verbalize_time for the voice agent? should it be a new tool for LLM or can I do it at the TTS layer? how will it impact the latency?

Thanks for reaching out, Shriniket. I'm an AI assistant and I'm here to help. Give me a moment while I research this topic.

I've already added in the prompt like this: 10:30 AM → "सुबह साढ़े दस बजे" but it doesn't work reliably, it works some time, but some time it goes like दस बज के तीस मिनट बजे which is not the right way of saying the time in hindi. You already get 10:30 AM from RAG so llm have that info and for TTS I'm using Cartesia

Thanks — I hear you: the LLM prompt mapping is producing inconsistent Hindi renderings (e.g., correct "सुबह साढ़े दस बजे" vs. awkward "दस बज के तीस मिनट बजे"). Since you already receive a precise time (10:30 AM) via RAG and you use Cartesia TTS, the most reliable approach is to convert the time into a deterministic, locale-correct Hindi string before TTS (TTS-layer or small preprocessing helper). This avoids LLM variability and keeps latency low. What to do (recommended) 1. Preprocess times into canonical Hindi phrases before sending text to Cartesia. ◦ Replace placeholders like

[TIME]

with a formatted string produced by a small, deterministic function. ◦ This can live in the TTS pipeline or a lightweight middleware right after RAG/LLM output. 1. Use explicit rules for common patterns (examples) ◦ Exact hour (10:00 AM) -> "सुबह दस बजे" ◦ Half hour (10:30 AM) -> "सुबह साढ़े दस बजे" ◦ Quarter past (10:15 AM) -> "सुबह सवा दस बजे" ◦ Quarter to (9:45 AM) -> "दस बजे से चौथे मिनट पहले" — but prefer "सवा चार" patterns for common usage (see below) ◦ Minutes 1–29 -> "X बजे Y मिनट" only when no idiomatic phrase exists ◦ Prefer idiomatic Hindi for halves/quarters and common colloquial forms: ▪︎ :15 -> "सवा [hour]" (e.g., 10:15 -> "सवा दस बजे" or "सवा दस") ▪︎ :30 -> "साढ़े [hour]" (e.g., 10:30 -> "साढ़े दस बजे") ▪︎ :45 -> "पौने [next_hour]" (e.g., 9:45 -> "पौने दस बजे") ◦ Use period-of-day prefixes: रात/सुबह/दोपहर/शाम based on hour ranges (e.g., 5–11 -> सुबह, 12–15 -> दोपहर, 16–19 -> शाम, 20–4 -> रात) 1. Example deterministic function (pseudo)

def verbalize_hindi(dt):
    h = dt.hour % 12 or 12
    m = dt.minute
    period = choose_period(dt.hour)  # returns 'सुबह','दोपहर','शाम','रात'
    if m == 0:
        return f"{period} {h} बजे"
    if m == 30:
        return f"{period} साढ़े {h} बजे"
    if m == 15:
        return f"{period} सवा {h} बजे"
    if m == 45:
        next_h = (h % 12) + 1
        return f"{period} पौने {next_h} बजे"
    return f"{period} {h} बजकर {m} मिनट"  # fallback

(Adjust wording to your preferred register and remove "बजे" if Cartesia voices prefer shorter phrases.) 1. Cartesia-specific notes • Cartesia supports SSML; use it to control prosody/pause around the time phrase for clearer speech. • Send the fully formed Hindi string to Cartesia (avoid leaving time-formatting to the LLM). • This keeps latency minimal: only your small formatting function + Cartesia TTS synthesis time. 1. Handling mixed cases • If you need context-aware phrasing (e.g., “in two hours” or business-day logic), compute that programmatically and pass the resolved absolute time into the same deterministic formatter before TTS. • If you must let the LLM craft the phrase, constrain it tightly with a function tool that returns the preformatted Hindi string (but expect higher latency). If you want, I can: • Provide a ready-to-copy implementation in your language (Python/Node) for Cartesia integration and SSML wrapping. • Add locale rules for formal vs. colloquial Hindi phrasing. Which would help you most: a full code example in Python or Node?