How to moderate voice messages, to the second

To moderate voice messages, send each recording to a speech moderation API that turns the speech into text with the time of every word, checks that text for the categories your rules forbid, and returns each finding with its start and end in seconds. A moderator then plays only the flagged seconds, and your app can hold, mute or deliver the message. Detectivision's audio moderation does this in one API call, in 12 languages.

Why voice messages need their own check

Chat apps, dating apps, games and communities that filter typed messages often let voice messages through unread. Abuse moves to where the filter does not look: a voice note in a match's chat, a clip in a game lobby, a recording in a group.

• Nobody can listen to everything. A team can read a flagged text in seconds. A voice message has to be played, and most are harmless.
• Reports come late. By the time someone reports a voice message, the person it was aimed at has heard it.
• The flagged part is short. One insult in a two-minute message is a few seconds long. Without its time, a moderator plays the whole recording to find it.

How speech moderation works

1. Speech becomes text. A speech model transcribes the recording and keeps the time of every word.
2. The text is checked by category. The transcript is checked for insults, hate speech, profanity and the other categories you ask for, the way a typed message is.
3. Each finding gets its time. The answer gives the category, how sure the check is, and the start and end of the words that raised it.

What makes it harder than text: accents and dialects, background noise and music, several people talking at once, two languages in one sentence, and very short clips. Test with recordings from your own users before you set your lines.

What to do with a finding

1. Decide per category. Hate speech may block the message at once; profanity may only add a warning; an insult may go to a moderator. Each category and each confidence band gets its own action.
2. Hold or deliver. For one-to-one messages, such as on a dating app, check before delivery and hold what is flagged. For group chats and recordings that stay online, deliver and check in the background, then remove what breaks a rule.
3. Play from the second. Show the moderator the flagged span and play from its start. On recordings that stay published, mute or cut only that span.
4. Look at the sender. Several flagged messages in a day are a pattern; one is a mistake. Keep the sender's id with each check.
5. Explain and allow appeals. Tell the sender which rule the message broke, and let them ask for a second look.

How Detectivision handles it

Send the address of the recording to POST /api/v1/moderation with "type": "audio_moderation". The speech is transcribed with the time of every word and checked for your flags: profanity, insult, hate, sexual, self_harm and bad_habits. Each finding comes back with its start and end in seconds from the beginning of the recording.

curl https://api.detectivision.ai/api/v1/moderation \
  -H "Api-Key: $DETECTIVISION_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "type": "audio_moderation",
    "input": "https://example.com/voice/msg_5521.mp3",
    "async": true,
    "metadata": { "senderId": "u_481", "chatId": "c_9920" }
  }'

{
  "id": 1071,
  "type": "audio_moderation",
  "mode": "capture",
  "status": "completed",
  "flagged": true,
  "flags": [
    {
      "flag": "insult",
      "confidence": 0.91,
      "detections": [
        {
          "label": "insult",
          "confidence": 0.91,
          "start": 12.4,
          "end": 13.8
        }
      ]
    }
  ],
  "checked": ["hate", "insult", "profanity", "sexual", "self_harm"],
  "unchecked": [],
  "tookMs": 2300,
  "usage": { "charged": 1, "remaining": 1999 },
  "metadata": { "senderId": "u_481", "chatId": "c_9920" }
}

Sent with "async": true, the API answers at once with the request's id and the result comes to your webhook, signed, with your metadata. Values in the example are illustrative.

Speech in a live stream is checked the same way as it comes in, next to the stream's frames.

Speech moderation APIs compared

Each company's facts are from its own pages, as read on 5 November 2026. Detectivision is not affiliated with any of these companies. Hive's docs list mp3, ogg, wav, m4a and flac among its audio formats; Amazon says its toxicity detection also uses tone and pitch, not only the words.

Start with your own recordings

Take voice messages your moderators have already judged, removed and approved, in each language your users speak. Send them through the dashboard's Playground or the API, and count, for each flag, the abuse it missed and the harmless messages it flagged. Then set a line per flag: act on its own above it, and send the band below it to a person who plays the flagged seconds.