Voice notes into text

Clients brief by voice note, in dialect, with English words mixed in. Transcribe every one, and check what the transcript can get wrong before you rely on it.

The engine [official guidance]: ElevenLabs Scribe v2, through its API or the ElevenLabs speech-to-text page. An independent benchmark found it the most robust on Egyptian–English and Saudi–English code-switched speech, about 13% word error overall. Judge mismatches word by word, never by that figure. Noisy phone notes are less certain than clean recordings.

Steps:

1. Collect every note as it was sent (.opus, .m4a, .mp3, video). Don't convert it.

2. Keyterms: the brand and product names, the client's English marketing terms, people and places, spelled exactly as the brand writes them (the API takes up to 1,000).

3. Transcribe with the language detected automatically (notes mix Arabic and English), the speakers labelled, and every word timed (settings in Do it yourself below).

4. Write one file beside the audio with three things: the Arabic transcript as returned, with speakers and timecodes; a clean English translation of its meaning; and a list of uncertain spans: names, numbers, prices, dates, any English term that came back in Arabic script, any low-confidence word.

5. Check the coverage on the transcript's timestamps (how to export them: below). The last word should end near the end of the audio. List any gap over 2 s with no words and listen to it: speech there means words were missed. Also check that every keyterm came back spelled as you gave it.

6. Check the uncertain spans (decided): Egyptian spans by your own ear when Egyptian is your own dialect, otherwise by a native Egyptian listener; in a Saudi or Gulf note, any span that changes a decision (a mandatory, a claim, a price, a name) is confirmed with the sender or the native reviewer before it enters the project file.

7. No voice-note line is quoted in a client document until it has been checked (decided).

Example. A note says:

«عايزين اللقطة الأخيرة على العلبة، مش على الموديل»

Translation: "They want the last shot on the box, not on the model." "Model" is an English word said in Arabic, so it is listed as uncertain until the translation settles it.

Cost: about $0.22 an hour of audio, plus $0.05 an hour for keyterms. Check the client's NDA before uploading their material to any outside service.

Do it yourself

Transcribe a voice note (ElevenLabs Scribe v2)

1. Upload the note as it was sent, on the ElevenLabs speech-to-text page or through the API.

2. Add your keyterms.

3. Settings: language automatic, speaker labels on, word-level timestamps.

4. Export the transcript with its timestamps: SRT (a subtitle file) or JSON (a data file) from the speech-to-text page, or word timestamps from the API.

Sources

• ElevenLabs, Scribe v2.