Back to blogClose-up of a microphone with a foam windscreen and black headphonesGEO Analytics

    Brand names in audio: transcription and disambiguation

    2026-11-03·7 min read

    Before counting an audio mention, check which name was actually spoken. If transcription replaces a word with your brand, the counter may record an appearance that was never in the recording. If it omits the name, a later literal search cannot recover it.

    This guide is for teams reviewing authorized recordings to evaluate recognition of brand names. The deliverable is an audio set with a human reference, transcription conditions and localized errors. It does not measure whether Siri or Alexa recommends a company, or turn transcription into evidence of endorsement.

    The voice assistant guide covers presence in those channels. This review addresses an earlier stage: correspondence between the sound and the name that reaches the text.

    Separate sound, spelling and identity

    Keep three decisions distinct: what is audible, how it is spelled and which entity it refers to. They may coincide, but the same evidence does not always resolve all three. An official spelling identifies a catalog name; it does not establish that the word was spoken in a particular recording.

    For each name, record approved spelling, accepted aliases and language. Document pronunciations present in your material without treating them as a universal list. Do not normalize a similar word to the brand merely because that is the entity you want to measure.

    When two names sound identical, isolated audio may not resolve identity. Preserve the ambiguity and the context that could clarify it. A recording owner's stated intention may help build a reference for a controlled test, but it does not make identical sounds distinguishable.

    Keep presence separate from evaluation too. A brand spoken in a negative statement remains an acoustic mention. Whether it is recommended requires another analysis of the content and its conditions.

    Build a human reference without revealing the prediction

    Use recordings you own or are authorized to send to the chosen service. Remove unnecessary personal data and out-of-scope material before sending them. Permission to listen to a call does not automatically authorize uploading it to any provider.

    Give each file an identifier, version, origin and known conditions. Keep original audio separate from derivatives. Record cropping, noise reduction and speed changes: those transformations can alter the evidence the system receives.

    Conduct an initial listen without showing the automatic transcript or expected-name list. Ask the reviewer to retain the audible passage and uncertainty rather than complete a mandatory brand label. Add any necessary identity information in a separate field afterward.

    Resolve disagreements through another independent listen and preserve initial decisions. If the reference remains ambiguous, do not force a positive or negative label for convenience. Identify which level is assessable: audible word, spelling or entity.

    Include controls where the brand is not spoken

    A set containing only mentions of your brand cannot reveal names introduced by the system. Include recordings with the brand, without it and with potentially confusable terms, all under a reviewed reference.

    Define the scope of absence: a bounded excerpt or the complete recording. Do not label an excerpt negative merely because it cuts out the mention. Silence and damaged files can serve as separate technical controls; they do not replace negative recordings containing real speech.

    Cover conditions relevant to the intended use: languages, accents, code-switching, microphones, noise and overlapping speakers. Do not infer those conditions from personal attributes or present a difficult-case battery as representative of all listeners.

    Group versions of the same audio and dependent recordings when separating development from held-out testing. Do not tune vocabulary on one excerpt and present a nearly identical crop as an independent evaluation.

    Compare context and vocabulary on the same audio

    Freeze a reference condition with the minimum necessary configuration. Compare it with another condition adding relevant context or vocabulary, using exactly the same files. Retain the service, exposed model or version, configured language and every submitted field.

    OpenAI's transcription documentation distinguishes free-form context, expected terms and languages. It states that keywords are hints rather than required output and should appear only when present in the audio. Check the options supported by your particular route; do not assume parity between an API and a commercial application.

    Do not instruct the system that the output must contain the brand. A long list of names is not neutral context either. Retain its contents and version so observed differences correspond to a documented condition.

    Google's Speech-to-Text adaptation guide warns that boosting terms can reduce omissions while increasing false positives. The comparison must therefore retain negatives rather than merely check that the expected name is now spelled correctly.

    Choose hints using the development set and freeze them before opening the held-out test. Do not repeatedly adjust context for a failed recording until the desired transcript appears, then call it a first-pass success.

    Preserve the original text before correcting it

    Store literal recognition output with the audio and condition. An editorial correction or entity normalization must create another field or version, not replace the output under evaluation.

    FieldWhat it makes reviewable
    Audio and intervalWhich excerpt was heard and its source version
    Human referenceAudible words, resolvable identity and ambiguity
    ConfigurationSubmitted route, language, context and vocabulary
    Original textRecognition output before corrections
    NormalizationWhat a later rule changed and why
    DecisionCorrect name, omission, substitution, insertion or pending

    Do not use the brand dictionary to turn every approximate match into a success. Decide in advance which spelling differences are equivalent for the task. Preserving identity may be acceptable for an indicator but not for literal transcription; show both results if both are needed.

    The AI video citation audit explains how to link tracks and intervals when the audio comes from a demonstration. Do not attribute a word to the recording when it appears only in captions or the video description.

    Count omissions and insertions with separate denominators

    Define whether the unit is a spoken name or an excerpt with brand presence. Do not mix them: an excerpt can repeat a name while remaining one presence unit. For this record, one option is excerpt plus target entity, with boundaries set beforehand.

    Among assessable positive-reference excerpts, count those where the system preserves the entity and those where it omits or substitutes it. Among assessable negatives, count those that introduce the brand. Publish counts and sizes for both sets; better positive results may accompany more insertions in negatives.

    Show ambiguous references, processing failures and incomplete outputs separately. Do not turn a timeout into absence or an uncertain reference into a recognizer error. For any rate, declare included cases and pending ones; a zero denominator provides no calculable rate.

    An overall word accuracy indicator can hide the name error that matters. Break results down by entity and acoustic condition, with available sample sizes. Do not treat a service's confidence score as certainty about identity without checking its relationship to observed successes and errors.

    Do not blame the classifier for a recognition error

    Retain the stages: audio, original transcript, normalization and mention or recommendation classification. Identify where the discrepancy first appears. If only the final output is observable, record that its cause cannot be localized.

    Mention classifier validation starts with available text. A correct label for a name inserted by transcription does not establish that the brand was in the audio. Likewise, a classifier cannot compensate for an omission without other evidence.

    If you test a later correction, report its result separately and retain any new errors it introduces. Do not credit the recognizer with the reviewer's successes or the classifier with a manual replacement rule's results.

    Use the same assessable excerpts when comparing conditions, while also preserving attempts that failed in either condition. The paired result and processing coverage answer different questions; neither should disappear because it makes comparison harder.

    Close the test with explicit limits of use

    The final record should contain set version, reference, configuration, counts by condition, localized discrepancies and the usage decision. Before testing, define which errors require human review and who accepts the limits; no universal threshold applies to every brand and recording.

    Keep confirmed false positives in regression alongside genuine mentions and new held-out material. Reassess affected cases after vocabulary, preprocessing or service changes. Improvement only on material used to tune hints is not independent validation.

    This protocol can accompany your Mentio visibility analysis. It does not assume that Mentio transcribes recordings or recognizes spoken names. The result is recognition evidence under specific conditions, not audience measurement or a guarantee that an assistant will recommend the brand.

    Frequently asked questions

    Does my brand appearing in a transcript prove it was spoken? No. It may be an insertion or substitution. Check the audio excerpt against a human reference and retain the original text, submitted context and review decision.

    Should I give the recognizer a list of brands? It can be a test condition when the route supports it and the terms are relevant. Compare it with a reference condition on the same audio and retain negative controls to detect added words that were not spoken.

    How do I resolve two brands that sound identical? Separate the audible form from identity. If sound and permitted context cannot distinguish the entity, retain an ambiguous label; official catalog spelling does not remove that acoustic limitation.

    Can I correct names before calculating the result? Preserve the unmodified transcript first. Evaluate normalization or correction as a separate stage and record its changes; do not credit the recognizer with a success produced afterward.

    Does recognizing the brand mean the assistant recommends it? No. Recognition checks correspondence between audio and name. Recommendation requires analyzing what is said about the entity, including negations, conditions and rejected alternatives.

    Want to know if AI mentions your brand?

    Discover your visibility in ChatGPT, Claude and Gemini in minutes.

    Related articles