Transcribing Oral-History Tapes: A Speech-to-Text Workflow for Accurate Names

Source Material: Preserve the Recording Before Transcription

An oral-history transcript is a representation of a recording, not a replacement for it. Begin by identifying the original carrier, recording format, duration, channel configuration and available playback equipment. A cassette, open-reel tape and digital audio file present different preservation requirements. Record the interview identifier, speaker names where known, recording date if documented and any existing restrictions on access.

Digitise analogue material through an appropriate playback and capture chain before attempting speech recognition. Retain a preservation master as an uncompressed or losslessly encoded audio file, with the original sample rate and bit depth documented. A WAV file using linear PCM is a common preservation choice. For example, 24-bit audio at 48 kHz provides a practical archival capture specification for many interview projects, although the source condition and institutional policy should determine the final settings.

• Keep the physical original and its existing documentation.
• Create a preservation master before making edited listening copies.
• Record the digitisation equipment and capture settings.
• Assign a stable identifier linking audio, transcript and metadata.

Condition: Identify the Difficulties That Affect Recognition

Speech recognition depends on more than the clarity of individual words. Background noise, tape hiss, room reverberation, overlapping speakers, regional accents and microphone distance can all affect the transcript. Older recordings may also contain dropouts, speed instability or sections where a speaker moves away from the microphone. These characteristics should be recorded before automated processing begins.

Make a distinction between defects in the recording and features of the speaker's delivery. Pauses, repetitions, dialect, hesitation and unusual pronunciation may carry historical or personal significance. Noise reduction can make a recording easier to hear, but aggressive processing may remove consonants or change the character of a voice. Preserve the original audio and make any listening enhancement a separate, documented derivative.

Intervention Log: From Automated Draft to Human Transcript

Use speech-to-text to establish a first draft, then organise correction around the recording rather than the software's apparent confidence. Divide a long interview into manageable passages and retain time references for each segment. A transcript with a timestamp every 30 seconds is a useful starting point; more frequent references are appropriate when speaker changes or difficult passages require precise location.

Correct obvious recognition errors while retaining the distinction between spoken words and editorial additions. Establish a consistent convention for interruptions, unintelligible speech, overlapping voices and non-verbal sounds. If a speaker repeats a word, the transcript should normally preserve that repetition. If an editor supplies a missing name from independent documentation, identify the intervention in the editorial policy rather than silently presenting it as something audible.

A practical correction sequence is to listen to the complete passage, inspect the proposed wording, replay the relevant seconds and enter the correction. For difficult names, repeat the process using a longer passage because surrounding context may clarify pronunciation. Keep a record of the transcription tool and relevant processing settings, but do not treat a confidence score as evidence that a word is historically correct.

• Retain the first automated draft as a separate working file.
• Correct against the source recording.
• Mark uncertain speech consistently.
• Preserve timestamps and speaker changes.
• Record substantive editorial additions.

Inference Disclosure: The Names That Matter

Proper nouns deserve a separate editorial pass. Speech recognition may turn an unfamiliar surname into a familiar word, merge two names or substitute a nearby place name. These errors can damage catalogue searches and make an interview difficult to connect with other historical records. A person mentioned for only a few seconds may be the key to identifying an event, photograph or location.

Create a name authority worksheet containing the transcript spelling, the audible pronunciation, the proposed spelling, the supporting source and the level of certainty. Consult interview notes, collection records, contemporary directories and other independently documented material where available. For example, if a speaker says a surname that sounds like 'Mackenzie', an electoral register may help establish the spelling, but the transcript should not imply that the register proves the recording's pronunciation.

Distinguish three editorial situations: a word clearly heard, a spelling supported by external evidence and a name that remains uncertain. Use a consistent notation for the third category, such as '[surname unclear, sounds like ...]'. Avoid filling a gap simply because a software suggestion looks plausible. If the evidence does not settle a spelling, retain the uncertainty and explain it in the transcript notes.

Reversibility and Outcome: Keep Every Representation Connected

A sustainable oral-history workflow produces several related but distinct objects: the original carrier, a preservation audio master, an access copy, the initial automated transcript, the corrected transcript and any supporting documentation. Each object should have its own identifier or filename linked through collection metadata. Never overwrite the preservation master with denoised audio or embed corrections in a way that destroys the earlier transcript.

Digital video preservation introduces a related requirement when interviews include synchronised picture and sound. Preserve the original audiovisual recording as a preservation master and maintain a documented relationship between its audio track, video stream and transcript. An audio-only extraction may be useful for speech recognition, but it is an access or working derivative. Record its relationship to the parent audiovisual file, including any changes to duration or synchronisation.

For long-term management, document file format, duration, checksum, storage location and creation date. A checksum can help identify accidental changes to a digital file, but it does not establish that the transcript is accurate. Maintain at least two independently managed copies of important preservation files, with routine fixity monitoring under the institution's preservation policy.

Questions readers ask