Text Polishing

Text Polishing removes spoken-language artifacts from a Scribe transcript and returns clear written text. It is optional and is available for en-US speech in Fast, Batch, and Live modes.

Text Polishing can:

• Remove filler words, false starts, and repetition
• Improve grammar, punctuation, capitalization, and readability
• Condense unnecessarily verbose speech
• Reorganize spoken thoughts for clarity
• Preserve names, facts, terminology, and intended meaning

Text Polishing can't be combined with config.language_detection, even when config.language is en-US. Set config.language to en-US and omit language_detection.

Text Polishing does not summarize the transcript or turn spoken instructions into a numbered or bulleted procedure.

Enable Text Polishing#

Add outputs.text_polishing as a top-level field alongside config.

{
    "file": "https://example.com/audio.mp3",
    "config": {
        "language": "en-US",
        "diarization": true
    },
    "outputs": {
        "text_polishing": {
            "enabled": true
        }
    }
}

If outputs.text_polishing is omitted or enabled is false, Scribe returns only the transcript.

Fast and Batch responses#

In Fast and Batch modes, each transcription returns one polished result. The original transcript remains under result and is never overwritten, and the complete polished text is returned under outputs.text_polishing.text. In Fast mode, the response is returned after both transcription and polishing are complete.

{
    "request_id": "3f2b8c1e-7a4d-4e9b-b6f0-1c5d8a2e9b47",
    "duration_sec": 27.4,
    "result": {
        "text_display": "Um, welcome everyone, uh, let's get started.",
        "segments": [
            {
                "id": "utt_0000",
                "start": 0.0,
                "end": 0.9,
                "channel": 0,
                "speaker": "Speaker 1",
                "text": "um, welcome"
            },
            {
                "id": "utt_0001",
                "start": 1.0,
                "end": 2.1,
                "channel": 0,
                "speaker": "Speaker 1",
                "text": "everyone,"
            },
            {
                "id": "utt_0002",
                "start": 3.5,
                "end": 6.2,
                "channel": 0,
                "speaker": "Speaker 2",
                "text": "uh, let's get started"
            }
        ]
    },
    "outputs": {
        "text_polishing": {
            "text": "[Speaker 1] Welcome everyone. [Speaker 2] Let's get started."
        }
    }
}

Batch uses the same outputs field. Each completed output file includes its own outputs.text_polishing.text value.

Live responses#

Enable Text Polishing in session.update.

{
    "type": "session.update",
    "config": {
        "language": "en-US",
        "diarization": true
    },
    "audio": {
        "format": "pcm16"
    },
    "outputs": {
        "text_polishing": {
            "enabled": true,
            "include_transcript": true
        }
    }
}

In Live mode, polished results are delivered in transcription.completed events. One polished event can cover several speech items: it uses the first item's item_id, its audio_start_ms and audio_end_ms span all of the covered items, and the later items don't receive their own transcription.completed event. Don't wait for a transcription.completed event for every item_id from speech_started. outputs.text_polishing.text contains the polished text for the items that event covers, so append each result to your document; a later event does not replace earlier text.

{
    "type": "transcription.completed",
    "item_id": "a1b2c3d4e5f6",
    "transcript": "[Speaker 1] um, welcome everyone, glad you could make it",
    "audio_start_ms": 0,
    "audio_end_ms": 4200,
    "transcription_latency_ms": 240,
    "outputs": {
        "text_polishing": {
            "text": "[Speaker 1] Welcome everyone. Glad you could make it."
        }
    }
}

With diarization enabled, Live polishing returns speaker labels as inline [Speaker N] tags in both transcript and the polished text, rather than in a separate speaker field.

include_transcript is specific to Live and defaults to false. When it is false, a completed event contains polished text without the original transcript. Polishing adds latency: Scribe sends transcription.completed only after polishing finishes.

Speaker and channel tags#

Polished text keeps speaker and channel labels as inline prefixes:

Consecutive sentences from the same speaker share one tag. A new tag starts when the speaker changes.

Handle polishing failures#

Transcription succeeds even when polishing fails. In that case, outputs.text_polishing.text contains only a failure marker. Fast still returns HTTP 200, a Batch file still completes as FILE_SUCCEEDED, and a Live session stays open.

Do not display a value that starts with [polishing_failed: as polished speech. In Fast and Batch, use result.text_display instead. In Live, use transcript, which is included only when include_transcript is true.

If the selected locale does not support Text Polishing, Fast and Batch return 400 INVALID_INPUT without a transcript. Live returns a nonfatal invalid_config error and does not apply the session update.

{
    "type": "error",
    "error": {
        "code": "invalid_config",
        "message": "Text polishing is not supported for locale 'zh-CN'",
        "fatal": false
    }
}