Transcriber (Welsh) for Conversational Audio

You will review automatic speech recognition drafts against Welsh conversational audio and correct each transcript word for word. The goal is a final word error rate of 5% or lower across a 1,350-hour voice dataset, delivered in 450-hour tranches.

Each file includes the audio, an editable ASR text draft, and word-level ASR JSON for reference.

• Listen to every file in full and preserve fillers, repetitions, stuttering, false starts, incomplete sentences, dialect, and backchannels.
• Keep English speech in English. Re-transcribe any segment where the ASR system changed English speech into Welsh.
• Label speakers and correct start and end timestamps to 0.01 seconds.
• Split segments at natural pauses and keep them under 30 seconds where possible.
• Deliver JSON segments containing the speaker, start time, end time, and transcript text.
• Do not add tags such as (laughter) or (cough).

What it pays and takes

This is a contractor role for 10 Welsh-speaking transcribers. The project is expected to run from mid-November through December 31, 2026, with a tentative start date of November 16.

• Pay: $30 to $35 per hour.
• Schedule: 8 hours per day, Monday through Friday.
• Time requirement: 20 or more hours per week.
• Language: Fluent Welsh, with the ability to handle Welsh conversational speech and English code-switching.
• Experience: Prior transcription experience is required.
• Preferred experience: Correcting ASR output and producing verbatim transcripts.
• Technical requirements: Comfort working with JSON and precise timestamping.
• Availability: Full-time availability from mid-November through December 31, 2026.

How it works

Apply on OpenTrain, where the employer reviews applications.

About AI training work

AI training work uses human-prepared examples, such as accurate transcripts, to help improve artificial intelligence systems. Careful language and audio reviewers are important because they help models handle real speech, accents, pauses, and mixed-language conversations.