Japanese (ja-JP) STT outputs kanji for spoken names — can it return a phonetic (katakana/kana) reading instead?

Support Help
1

Hi Retell team,

We’re building a Japanese-language voice agent (Conversation Flow) that collects
the caller’s name during a reservation and reads it back for confirmation.

Problem

For a Japanese voice agent, the STT returns spoken personal names as KANJI
(natural Japanese orthography), not as the phonetic reading the caller actually
spoke. Kanji is ambiguous about its reading, so the spoken information is lost the
moment the STT commits to a kanji spelling.

Concrete example from a real call:

  • The caller clearly says the reading “Toyoshima” (とよしま).
  • The STT transcript comes back as the kanji “豊島”.
  • “豊島” has multiple valid readings (Toshima OR Toyoshima), and we have no way to
    recover from the kanji alone which one the caller actually said.
    (Other everyday examples: 東 = Higashi or Azuma; 新谷 = Shintani or Araya; etc.)

To be clear about where the conversion happens

The kanji-to-kana conversion is done on OUR side, downstream of Retell — we run our
own LLM to guess a reading from the kanji STT output. We are NOT asking Retell to
do kana conversion. We only do this lossy guess because the STT gives us kanji in
the first place; when it guesses the more common reading (“Toshima”), the agent
reads back the wrong name and the caller has to correct us.

If the STT could return the phonetic reading directly, we could drop this lossy
downstream step entirely.

Technical details

  • Agent: agent_57b261c7d9fdd102163a5f0fd0 (name: voiceai_stg_a3867e4c)
  • Response engine: Conversation Flow (type “conversation-flow”, version 0)
  • Channel: voice
  • Languages (multilingual agent): ja-JP, en-US, ko-KR, zh-CN
  • Voice (TTS): voice_id custom_voice_c9cd94fd5a53bca14cc385c189, voice_model “s2-pro”
  • STT-related agent settings: only language and boosted_keywords
    ([“さくらカフェ”,“ランチ”,“洋食”,“バイキング”,“ヒスイ”,“中華”,“ディナー”]) are set.
    stt_mode, vocab_specialization, denoising_mode, normalize_for_speech,
    and pronunciation_dictionary are all unset (defaults).
  • Example call exhibiting the issue: call_e5c7b48dcf661f1112911876eef

Questions (all about the STT output, not downstream)

  1. Can the Japanese (ja-JP) STT be configured to return a phonetic reading
    (katakana / hiragana / furigana) for recognized speech, instead of — or
    alongside — the kanji?
  2. Does the API expose an alternative or phonetic transcript / word-level reading
    for a recognized segment that we could use for proper nouns?
  3. Are there any STT settings (model selection, stt_mode, vocab_specialization,
    formatting options, custom vocabulary) that bias the Japanese transcription
    toward phonetic output, particularly for name-collection steps?
  4. Which STT provider/model is used for Japanese (ja-JP) on a Conversation Flow
    voice agent, and does it support a reading/furigana output mode?

The reading — not the kanji — is what we need for an accurate spoken read-back and
for our records, so any way to obtain the phonetic form at the STT layer would
solve this for us.

Thanks in advance for any guidance.

2

Thanks for the detail write up , i have escalated your Call ID and Questions to the team for a review. Regards

3

Unfortunately no, Retell’s Japanese STT doesn’t have an option to force phonetic output instead of kanji. Your best bet is using the pronunciation_dictionary setting in agent config — you can map kanji names to their kana readings, and the STT should respect those hints. If that doesn’t work reliably, you might need to ask callers to spell their name phonetically or switch STT providers that support this.

4

Hi @rweale

Thank you for your answer!! I’ll find another solution.

Goro

5

Hey @goro.yanagi We currently have no option for katakana output; the transcript is delivered directly from the STT provider. We’re continuing to investigate possible solutions.