Transcribing Recorded Speech
Chapter Objectives
By the end of this chapter, you will be able to:
- Define the recording, the language variety, and the transcription purpose before beginning any transcription task.
- Segment and replay audio systematically rather than transcribing on a single uncontrolled pass.
- Draft, test, and revise a transcription against the recording itself.
- Use slowed playback and visual evidence such as a waveform or spectrogram responsibly, without treating either as an automatic transcription.
- Document uncertainty in working notes rather than concealing it behind a confident final form.
- Distinguish what you actually heard from what spelling or a remembered word led you to expect.
- Complete a short transcription that is fully supported by a verified recording.
An Observable Problem
Chapter 26 handed you a transcription that someone else had already produced and asked you to read it accurately. This chapter reverses that task one more time. You now have a recording, no transcription, and a job to do: listen, decide what was said and how, and write it down using IPA that another trained reader could rely on.
This is a harder task than reading, and a different one. A finished transcription on a page is fixed; a recording is not. The same stretch of speech can be replayed at full speed, replayed slowed down, viewed as a waveform, or viewed as a spectrogram, and each pass can reveal something the others missed or obscured. A transcriber who listens once and writes down a confident guess is not transcribing. A transcriber who works through a controlled, repeatable procedure, checks each decision against the recording, and documents what remains uncertain is doing the job this chapter teaches.
This chapter asks three questions:
- What sequence of steps turns a recording into a defensible transcription, rather than a first impression written down and left unchecked?
- How should slowed playback, waveforms, and spectrograms support a transcriber's ear without replacing it?
- How does a transcriber keep hearing separate from expectation, so that a transcription reflects the recording rather than a guess about the underlying orthography?
A Nine-Step Transcription Workflow
Purpose: replace a single uncontrolled listen-and-write pass with a disciplined procedure that produces a transcription you can defend.
Chapter 26 gave you a seven-step strategy for reading a transcription that already existed. Transcribing a recording adds earlier steps, because you must first establish what you are transcribing and why before any symbol is written, and later steps, because a draft transcription must be tested against the recording rather than accepted on the first attempt.
Listening globally first, before any transcribing begins, prevents a common failure: fixating on the first word or syllable and building an increasingly strained analysis around an early misperception. A global pass establishes the overall prosodic shape, likely word or phrase boundaries, and general register of the recording, so that later, detailed listening happens with realistic expectations already in place.
Establishing Metadata and Task Level
Purpose: apply the reference-variety and transcription-purpose policies from Chapters 24 and 26 at the start of a recording-based task, before any listening for content begins.
A transcription produced without stated metadata cannot be evaluated. Before transcribing any recording, establish, as far as the available information allows:
| Item | Why it matters |
|---|---|
| Language and reference variety | Determines which phoneme inventory and realization patterns are expected, per Chapter 10's reference-variety policy. |
| Speaker context | A single speaker, multiple speakers, a scripted reading, or spontaneous speech each carry different expectations for consistency and connected-speech processes. |
| Requested transcription level | Phonemic, broad phonetic, or narrow phonetic, following the Chapter 24 distinction; this determines how much detail step six should add. |
| Source status | Whether the recording is an idealized citation-form reading or a naturally occurring utterance affects which connected-speech processes from Chapter 23 you should expect to hear. |
Skipping this step does not save time. A transcriber who begins listening for content before settling the task level typically produces an inconsistent transcription, mixing phonemic and narrow phonetic detail without a principled reason, the exact problem Chapter 24 identified as a common transcription error.
Listening Globally Before Segmenting
Purpose: connect the global-then-detailed listening sequence to the auditory training principles established across the book, and demonstrate it with a fully described worked example.
A global first listen is not casual listening. It is a directed pass with specific goals: counting apparent words or phrases, noticing overall pitch movement, noting where pauses fall, and forming a provisional sense of speaker style and rate. None of these goals require transcribing a single segment yet.
Consider a short, fully described recording used here as a worked demonstration. The recording contains one speaker of a named variety producing a two-word phrase at a conversational rate, with a brief pause between the words and falling pitch at the end. A global listen establishes this much before any symbol is written: two prosodic units, a boundary between them, and a terminal fall consistent with a completed statement rather than a question.
A global listen can reliably establish the number of prosodic units, the rough location of major boundaries, and the general pitch contour. It cannot reliably establish fine segmental detail, exact vowel quality, or the presence of subtle allophonic variation. Treat the global pass as a map of the territory, not as a transcription in miniature.
Drafting, Testing, and Revising
Purpose: teach the iterative core of the workflow, steps five through seven, as a cycle rather than a single pass.
A draft transcription is a hypothesis, not a final answer. After segmenting the recording and identifying syllables and prosody, write a broad transcription of each segment based on what you hear. Then immediately test that draft: replay the same stretch of recording while following your draft symbol by symbol, and ask whether each symbol still matches what is audible on a second and third listen.
This cycle regularly produces revisions. A segment that sounded like a plain voiceless stop on a first pass may reveal audible aspiration on a closer, repeated listen; a vowel transcribed as central on a first attempt may sound more front once compared directly against a neighboring, clearly front vowel elsewhere in the same recording. Revision at this stage is normal disciplined practice, not a sign of an unreliable ear.
Note that the finished, learner-facing answer for a task like this should still present one transcription as the submitted response, with the alternative and its supporting reasoning placed in the accompanying commentary, exactly as Chapter 24 required for two-level transcription work.
Using Slowed Playback and Visual Evidence Responsibly
Purpose: extend Chapter 28's caution about acoustic displays to the transcription workflow itself, and prevent a visual display from silently replacing auditory judgment.
Slowed playback and visual displays such as waveforms and spectrograms are legitimate, useful tools at step seven. Slowed playback can separate segments that blur together at conversational rate. A waveform can help locate a closure or a burst precisely in time. A spectrogram can help confirm frication, distinguish a stop from an affricate, or estimate approximate formant values, applying the introductory acoustic skills from Chapter 28.
These tools support a transcriber's ear; they do not replace it. Slowed playback can distort natural coarticulation and voice quality, so a decision made only from slowed audio should be re-checked at natural rate before it is finalized. A spectrogram shows acoustic patterns, not a finished transcription; reading a spectrogram as if it printed IPA symbols directly repeats the error Chapter 28 explicitly warned against.
Error: Basing a transcription decision entirely on a spectrogram image without confirming it against what is actually audible at natural speaking rate.
Correction: Use the spectrogram to support or question an auditory impression, then return to the recording at natural rate to confirm the decision. A transcription is a claim about what a listener can hear, not only about what a display can show.
Documenting Uncertainty
Purpose: apply the evidence-before-confidence principle to a transcriber's own working notes, and separate the honestly uncertain from the confidently supported.
Not every segment in a recording is equally clear. Background noise, overlapping speech, fast connected-speech reduction, or a genuinely ambiguous articulation can all leave a transcriber with real uncertainty. The workflow's eighth step exists specifically to hold that uncertainty rather than hide it.
Working notes should distinguish at least three levels of confidence: segments the transcriber is confident about and can defend from repeated listening; segments where two or more transcriptions are defensible and both should be recorded with their supporting evidence; and segments where the recording does not provide enough evidence to transcribe responsibly at all, which should be marked as such rather than filled in with a guess.
A transcription that marks an unclear stretch as uncertain, using a standard convention agreed with the project or instructor, is more useful and more accurate than one that fills the gap with a plausible-sounding guess. The goal of this chapter is a transcription whose confidence matches its evidence, not a transcription that looks complete at the cost of being unsupported.
Hearing Versus Orthographic Expectation
Purpose: extend the Chapter 26 reading-versus-guessing check to the act of transcribing from audio, where the same failure can occur in the opposite direction.
Chapter 26 warned against reading a transcription by guessing at a remembered spelling instead of interpreting the symbols present. Transcribing from a recording invites a related but distinct failure: hearing what a known spelling or a recognized word predicts, rather than what the recording actually contains.
This failure is strongest with familiar-sounding words, loanwords, or recordings in a language that shares vocabulary with the transcriber's own. A transcriber who recognizes a word before finishing the segmenting step may unconsciously transcribe the citation form they already know, rather than the specific realization actually present in that recording, including any reduction, assimilation, or speaker-specific variation.
Guard against this by completing steps two through five before consulting any gloss, translation, or orthographic form, exactly as the cover-the-gloss check in Chapter 26 recommended for reading. If a written form must be consulted for the task, treat it as a check on the finished draft, never as the source of the transcription itself.
Error: Transcribing a recognized word using its expected citation-form pronunciation, rather than the reduced or contextually altered form actually audible in the recording.
Correction: Transcribe what is audible on repeated listening first, and compare the result with the citation form only afterward, treating any difference as potential evidence of a genuine connected-speech process rather than as an error to silently correct.
Transcribing Across Recording Conditions
The nine-step workflow applies regardless of language, but recording conditions change how demanding each step is. A clear, single-speaker citation-form recording may move quickly through steps five through seven with few revisions. A multi-speaker recording of spontaneous connected speech, or a recording of a language with a large tonal or consonant-cluster inventory, will typically require more revision cycles at step seven and a longer list of documented alternatives at step eight. The steps themselves do not change; the amount of iteration each one requires does.
Analysis Exercises
These exercises apply the nine-step workflow to fully described, already-verified recordings rather than to unavailable audio, so that the reasoning process can be practiced honestly. Treat each item as a worked case: the relevant listening evidence is given in prose, and your task is to apply the workflow to it.
Exercise A — Metadata First
You are given a recording described only as "a greeting." Before listening for content, list the four metadata items from this chapter that must be established, and explain what could go wrong if each one were skipped.
Exercise B — Draft, Test, Revise
A transcriber's first-pass draft of a two-syllable word is [peto]. On a second, focused listen to the medial consonant, the transcriber notices a brief period of voicelessness at the very start of the following vowel that was not audible on the first pass. Using steps seven and eight, describe what the transcriber should do next and what the working notes should record.
Exercise C — Hearing Versus Expectation
A recording contains a word that closely resembles an English word familiar to the transcriber. Describe the specific sequence of steps the transcriber should follow to avoid transcribing the expected English-influenced pronunciation instead of the form actually present in the recording.
Common Mistakes
Error: Beginning detailed segmental transcription on the very first listen, before a global pass has established overall prosodic shape.
Correction: Complete a global listen first, noting prosodic units, boundaries, and overall contour, and only then move into segmenting and detailed drafting.
Error: Treating a first draft as final without replaying it against the recording.
Correction: Always complete the compare-against-the-recording step before submitting a transcription, and expect that step to produce revisions.
Error: Presenting an uncertain segment with unwarranted confidence rather than documenting the uncertainty.
Correction: Use the working-notes convention for uncertain segments, and record the specific evidence for and against each defensible alternative.
Chapter Review
Part A — The Workflow
1. List the nine steps of the transcription workflow introduced in this chapter, in order.
2. Explain why establishing metadata and task level comes before listening for content.
Part B — Evidence and Revision
1. Explain why a draft transcription should be treated as a hypothesis rather than a final answer.
2. Explain the responsible relationship between slowed playback, spectrogram evidence, and a transcriber's own ear.
Part C — Uncertainty and Expectation
1. Describe the three levels of confidence a transcriber's working notes should distinguish.
2. Explain how transcribing from expectation differs from the reading-versus-guessing failure described in Chapter 26, and how the same underlying discipline prevents both.
Cumulative Review
1. From Chapter 24: Explain how the requested transcription level, established at step one of this chapter's workflow, determines how much detail is added at step six.
2. From Chapter 28: Explain how introductory acoustic evidence supports step seven of this chapter's workflow without replacing auditory judgment.
3. Preview for Chapter 28: This chapter used waveform and spectrogram evidence only as a support to listening. Predict what additional acoustic concepts Chapter 28 will need to introduce before that evidence can be interpreted on its own terms.
Self-Check
Before moving to Chapter 28, confirm that you can do the following without looking at your notes:
- State the nine steps of the recording-based transcription workflow in order.
- Explain why metadata and task level must be established before listening for content.
- Describe how to draft, test, and revise a transcription against a recording.
- Explain the responsible, supporting role of slowed playback and visual displays.
- Document uncertainty using at least three distinct levels of confidence.
- Explain how expectation-driven mishearing differs from accurate transcription, and how to guard against it.
If any of these tasks are difficult, review the relevant section before reading Chapter 28. Chapter 28 introduces the acoustic concepts, such as waveforms, spectrograms, duration, fundamental frequency, and formants, needed to interpret the visual evidence this chapter introduced only as a support to listening.
Answer Key (click to expand)
Analysis Exercises — Exercise A
The four metadata items are language and reference variety, speaker context, requested transcription level, and source status. Skipping the language and variety risks applying the wrong phoneme inventory or realization expectations. Skipping speaker context risks misjudging which variation is speaker-specific rather than systematic. Skipping the requested level risks producing an inconsistent mix of phonemic and narrow phonetic detail. Skipping source status risks mistaking a natural connected-speech process for an error, or the reverse.
Analysis Exercises — Exercise B
At step seven, the transcriber should replay the medial consonant and following vowel repeatedly to confirm whether the brief voicelessness is consistently audible rather than a one-time impression. If confirmed, the draft should be revised to reflect a voiceless or partially devoiced realization, for example using an appropriate diacritic if the task level and evidence support that level of detail. At step eight, the working notes should record both the original and revised forms, the specific evidence for the revision, and a confidence level, since a single brief cue may warrant a moderate rather than a certain confidence rating.
Analysis Exercises — Exercise C
The transcriber should complete steps two through five, global listening, segmenting, identifying syllables and prosody, and drafting, entirely from the recording, before consulting the resemblance to the English word or any written gloss. Only after producing an independent draft should the transcriber compare it with the expected form, treating any difference as potential evidence of a genuine phonetic property of the recording rather than as an error to silently correct toward the more familiar English-influenced pronunciation.
Chapter Review — Part A
1. Establish metadata and task level; listen globally; segment the utterance; identify syllables and prosody; draft broad segmental content; add supported detail; compare against the recording; document alternatives; complete a consistency audit. 2. Metadata and task level determine which phoneme inventory, variety-specific expectations, and level of detail apply to every later decision; establishing them first prevents the transcriber from making detailed decisions that must later be discarded or revised because the basic scope of the task was unclear.
Chapter Review — Part B
1. A draft is based on a limited number of listens and represents the transcriber's best current interpretation, not a verified final claim; treating it as a hypothesis keeps the transcriber open to revision once it is tested against further listening. 2. Slowed playback and spectrogram evidence should confirm, question, or refine an auditory impression, and any decision reached primarily from a visual display or slowed audio should be checked again against the recording at natural rate before it is finalized; the transcriber's ear remains the primary instrument.
Chapter Review — Part C
1. Confident and defensible from repeated listening; genuinely ambiguous with two or more defensible alternatives recorded along with their evidence; and insufficiently supported by the recording to transcribe responsibly, which should be marked rather than guessed. 2. Chapter 26's failure involved substituting a guessed spelling for genuine interpretation of a transcription already on the page; this chapter's failure involves substituting an expected pronunciation for genuine listening to a recording. Both are prevented by the same discipline: complete the analysis from the primary evidence, IPA symbols in one case and the recording in the other, before consulting any secondary information such as a gloss or a familiar spelling.
Cumulative Review
1. Chapter 24 established that phonemic, broad phonetic, and narrow phonetic transcription serve different purposes and should not be mixed without reason; the task level fixed at step one of this chapter's workflow tells the transcriber exactly how much phonetic detail step six should responsibly add, preventing an arbitrary or inconsistent level of detail. 2. Chapter 28 will teach waveform and spectrogram interpretation as supporting evidence for claims about voicing, closure, frication, and formant structure; in this chapter, that evidence supports step seven's comparison against the recording without replacing the transcriber's own listening, consistent with Chapter 28's caution against treating a spectrogram as an automatic transcription. 3. Chapter 28 will likely need to introduce the basic organization of a waveform and spectrogram display, including time and frequency axes, periodic and aperiodic energy, and the visual correlates of voicing, closure, release, and frication, before the visual evidence used only informally in this chapter can be interpreted with real precision.
Looking Ahead
What You Can Now Do
You can now define the recording, variety, and transcription purpose before beginning a transcription task; segment and replay audio systematically; draft, test, and revise a transcription against the recording itself; use slowed playback and visual evidence responsibly, without letting either replace your own listening; document uncertainty honestly; and distinguish what you actually heard from what a familiar spelling or word led you to expect. Together with Chapters 24 through 26, you now understand how to decide what a transcription needs, how to notate it correctly, how to read someone else's finished transcription, and how to produce your own transcription directly from recorded speech.