1. Confirm that the participant's release form explicitly covers the production of a written transcript in addition to the audio or video recording. If the original consent does not mention transcription, contact the participant or their legal representative to obtain a supplementary authorisation;
2. Verify whether the transcript will be stored, shared, or published differently from the recording itself — if so, document the additional legal basis under GDPR; 
3. For pre-existing archival collections, apply the diligent-search procedure before transcribing materials without explicit release;
4. Identify any portions of the recording that must be redacted or anonymised in the transcript (names of third parties, sensitive personal data, information covered by agreed restrictions).

1. Listen to a representative sample of the recording before committing to automatic transcription;
2. Evaluate: overall audio quality and signal-to-noise ratio; number of speakers and clarity of turn-taking; presence of overlapping speech, strong accents, code-switching, or domain-specific terminology; recording language(s) and script;
3. Decide whether automatic transcription is appropriate. If audio quality is very poor, or if the recording features non-standard language varieties not well covered by available models, plan for manual or semi-manual transcription instead;
4. Note the total duration of the recording to estimate processing time and post-correction workload.


A. Convert and normalise the audio file
1. Convert the recording to a format accepted by the chosen ASR tool. For example, Whisper accepts MP3, MP4, MPEG, MPGA, M4A, WAV, and WEBM, but the recommended input format is WAV (16 kHz, 16-bit, mono) for optimal recognition accuracy.
You may use FFmpeg to extract audio from video files and convert or resample recordings accordingly. Never overwrite or alter the preservation master; always work on a derivative copy and document the full conversion chain.

B. Clean the audio signal (if necessary)
1. If the sample assessment revealed significant background noise, hum, or interference, apply noise reduction before transcription. Excessive noise substantially degrades ASR accuracy;
You may use Audacity to clean the recording before transcription by selecting a noise-only segment to generate a noise profile, then applying noise reduction to the full recording with conservative settings to minimise artefacts in the speech signal;
3. If the recording has very low volume, you may normalise the amplitude using Audacity’s normalization function or an equivalent loudness normalisation filter. Save the cleaned working copy in the same format as the conversion output (WAV 16 kHz, 16-bit, mono) and document all processing operations, including software version and parameter values;
16.	Save the cleaned working copy in the same format as the conversion output (WAV 16 kHz, 16-bit, mono). Document all processing operations applied, including software version and parameter values.

1. Choose the ASR tool according to your computational resources, required languages, and accuracy needs. The primary recommended tool is Whisper (OpenAI, MIT licence);
2. Set the source language explicitly using the language parameter. Avoid relying on automatic language detection for short recordings or recordings with code-switching: detection errors propagate through the entire transcript;
3. Select output formats according to your downstream needs;
4.  If the recording involves more than one speaker, run speaker diarisation to attribute each speech segment to a distinct speaker label before or alongside transcription. For recordings with only two speakers (interviewer / interviewee), verify diarisation accuracy: the model may misattribute overlapping speech or hesitation sounds. Correct misattributions systematically;
5. If diarisation tools are not available, annotate speaker turns manually (e.g. in ELAN or oTranscribe using the timestamped output from Whisper as a reference skeleton).

1. Establish and document transcription conventions before post-correction begins or consult the transcription conventions established for your project or institution. Conventions must cover: notation of overlapping speech, pauses, non-verbal sounds, unintelligible passages, and anonymisation placeholders. Consistent conventions are essential for corpus reuse;
2. Open the TXT or SRT output alongside the audio in a transcription interface that allows simultaneous playback and editing. Work through the transcript segment by segment, listening to each audio passage and correcting the text. Do not edit from the text alone without listening: errors that appear plausible in text are often audible in the audio;
3. Identify systematic error patterns: recurring misrecognitions of proper nouns, place names, technical terms, or dialectal forms; hallucinations (words or phrases inserted by the model with no acoustic basis); timestamp drift in long recordings;
4. Mark unintelligible passages with a standardised notation (e.g. [inaudible] or [unclear]) rather than guessing. Note the timestamp of the passage;
5. Mark non-verbal sounds relevant to meaning (laughter, crying, significant pauses, interruptions) using agreed conventions;
6. Where anonymisation is required, replace names and identifying information with agreed placeholders (e.g. [NAME], [PLACE]) consistently throughout the transcript. Record the mapping in a separate, access-restricted key document;
5. After textual correction, verify that the timestamps in the SRT or VTT output remain accurately synchronised with the audio. If significant drift is detected, re-run the alignment step;
6. Export the final validated version in all required output formats before proceeding to encoding.

N. B. If needed, you may calculate or estimate the Word Error Rate (WER) on a representative sample (200–300 words from different sections) by comparing the ASR output to a manually transcribed reference. WER below 10% indicates good ASR performance; above 20% signals that manual correction will be time-intensive and that the automatic output should be treated as a rough draft only.

1. For archival and interoperability purposes, encode the corrected transcript in TEI-XML using the TEI P5 guidelines, module 8 (Transcriptions of Speech). TEI-XML is the standard encoding format for primary textual sources in SSH digital humanities and is required by many European infrastructure repositories. Minimum TEI encoding for an oral source transcript includes: with full bibliographic and administrative metadata (interviewer, interviewee, date, location, project, language, transcription date, transcriber identity, encoding conventions); with ++(utterance) elements, who attribute for speaker attribution, and xml:id for segment identification; or @when attributes for timestamp anchoring; 
2. If full TEI encoding exceeds available resources, produce at minimum a well-structured plain text or SRT file with a separate metadata record in Dublin Core format; 3. Validate the TEI file against the TEI P5 schema using oXygen XML Editor or the TEI online validator before deposit.

Provenance documentation is essential for the reusability and citability of transcripts in subsequent research. An undocumented transcript cannot be responsibly cited or reused. 

1. Create a transcription documentation record for each transcript produced. This record is distinct from the field recording sheet (which documents the oral source itself) and must cover: the ASR tool used (name, version, model size); the input audio file name and format; all preprocessing operations applied (conversion, noise reduction) with software version and parameters; the language setting and task parameter; the output formats generated; the date of transcription run; the name of the corrector and the date of post-correction; the estimated or measured WER before correction; the transcription convention schema applied; any anonymisation applied and the location of the key document; 
2. Store this record as a structured metadata file (JSON or XML) alongside the transcript files. Link it to the field recording sheet for the corresponding oral source. 

Giorgia Lo Nigro

This workflow guides you through the automatic transcription of oral source recordings in SSH research contexts. It covers the full pipeline: audio preparation, model selection and configuration, transcription, speaker diarisation, quality control, manual post-correction, and the export of timestamped, structured transcripts suitable for archival deposit and open publication.
The central tool is Whisper, a multilingual, open-source automatic speech recognition (ASR) model developed by OpenAI and released under the MIT licence. The workflow also integrates complementary tools for audio preprocessing (Audacity, FFmpeg), enhanced transcription (WhisperX, faster-whisper), annotation (ELAN), and text encoding (TEI-XML). It applies equally to newly recorded interviews and to the retrospective transcription of archival holdings.

Edit actor

Other URLs relevant to the resource, e.g. ORCID, etc.

URL, ideally landing page, of the resource.

Person/group/body who helped create the resource.

Should be concise and raise interest in the entry. Allows <Link>Markdown</Link> content.

Other URLs relevant to the resource, e.g. code, DOI, etc.

Upload media

URL pointing to info about the access policy.

The activities you can do with the resource. See <Link></Link>.

Name of the conference where paper was given.

Describes the discipline covered by resource. See <Link></Link>.

Locations where the Resource is offered. See <Link></Link>.

URL to helpdesk for incidents & user requests.

Audience targeted by the resource. See <Link></Link>.

A particular published issue of a journal.

Concept or term related to MP entry. See <Link></Link>.

Language(s) in which a resource is available. See <Link></Link>.

Select license pertaining to resource. See <Link></Link>. If license is unknown, use the "Terms of use" property field to describe the usage terms.

Status of the Resource life-cycle. See <Link></Link>.

Mode of use for the resource. See <Link></Link>.

File format of the linked resource. See <Link></Link>.

Page #s of a resource, separate by commas.

Refine the resource's nature and function. See <Link></Link>.

Links to non-MP materials that are relevant.

Info about performance levels provider expected to deliver.

Standard used by the resource. See <Link></Link>.

Technology Readiness Level of the Resource. See <Link></Link>.

Resource user manual and documentation URL.

Version of the tool or service that is linked.

The volume of a journal or multivolume book.

Edit source

Name

If you are having trouble with the sign in process, please <ContactLink>contact the SSH Open Marketplace team</ContactLink>.

If you are having trouble with the sign up process, please <ContactLink>contact the SSH Open Marketplace team</ContactLink>.

Automatic Transcription of Oral Sources

Related items(12)

Introduction to Speech Analysis

Automatic Transcription of Dutch Speech Recordings (MP3 file)

Automatic Transcription of Dutch Speech Recordings (Ogg file)

Automatic Transcription of Dutch Speech Recordings (Wav file)

Workflow steps(7)

1 Verify legal authorisation for transcription

2 Assess the recording for transcription readiness

3 Prepare Audio

4 Select and configure the ASR tool

5 Assess and correct the transcript

6 Encode the transcript in TEI-XML (optional. Recommended for archival deposit)

7 Complete the transcription documentation record