Can Whisper Hallucinate Transcriptions? Why AI Transcription Makes Things Up

You’ve just transcribed a critical patient consultation, legal deposition, or hours of research interviews—only to discover that the AI transcription tool inserted words, phrases, or entire sentences that were never spoken. This is a documented failure mode known as transcription hallucination, and it can create serious problems when transcripts are used without verification.

In one 2024 study of 13,140 short English-language audio segments, researchers identified hallucinated phrases or sentences in 1.4% of the segments processed under the study’s test conditions. The rate should not be treated as universal: results can vary substantially with the model version, audio, language, speaker population, segmentation, and decoding settings. However, the study demonstrates why teams relying on automated transcription for high-stakes work need reliable review processes and access to the original recording.

The implications extend beyond ordinary transcription errors. Researchers found examples involving violent language, inaccurate racial associations, fabricated personal information, and nonexistent medical treatments.

Key Takeaways

In one study, Whisper produced hallucinated phrases or sentences in 1.4% of 13,140 short audio segments tested
Researchers classified 38% of the identified hallucinations as containing explicit harms, including violent language, inaccurate associations, made-up personal information, or false authority
A follow-up comparison found that five other speech-to-text services did not reproduce comparable hallucinated passages on the 187 samples tested, although that result does not prove those services never hallucinate
Silence, long pauses, background sounds, and other non-speech intervals are frequently associated with Whisper hallucinations
In the study dataset, hallucinations appeared more often in segments from speakers with aphasia than in control segments, raising accessibility and fairness concerns
Voice Activity Detection and other non-speech filtering methods can reduce hallucinations in some Whisper workflows, but their effectiveness depends on the model, audio, and configuration
Original recordings should remain available until high-stakes transcripts have been reviewed and approved
No automated transcription system should be assumed to be error-free, particularly in healthcare, legal, employment, or research settings

Understanding AI Transcription: How Speech Becomes Text

Modern AI transcription relies on neural networks trained on large audio datasets to convert spoken language into written text. The technology has transformed how professionals handle audio content by turning recordings into searchable text much faster than manual transcription.

Different systems use different architectures, but the process commonly includes several stages:

Audio processing: Raw audio is converted into a representation such as a spectrogram
Speech recognition: A model identifies patterns in the audio and maps them to text tokens
Sequence decoding: The system selects a probable sequence of words based on both the audio and patterns learned during training
Post-processing: Punctuation, timestamps, speaker labels, and formatting may be added

These systems can perform well on clear audio, but accuracy depends on recording quality, accents, overlapping speech, technical terminology, background noise, and how closely the audio resembles the model’s training data.

What Are AI Hallucinations in Transcription?

AI hallucinations in transcription occur when a speech-to-text system generates text that has no meaningful basis in the source audio. This is different from an ordinary substitution, such as transcribing “cat” as “hat.” A hallucination may introduce an entire phrase or sentence that was never spoken.

Researchers analyzing 13,140 audio segments identified 187 hallucinations. They classified 38% of those hallucinated passages as containing explicit harms, including:

Violent language: Inserting violent statements into otherwise innocuous speech
Inaccurate associations: Adding characteristics such as race that the speaker never mentioned
Medical fabrications: Inventing nonexistent treatments or medication names
False authority: Adding language that resembles a quotation, caption, or authoritative statement

These examples are concerning because fabricated text can appear fluent and credible, making it difficult to notice without comparing the transcript with the source audio.

A separate 2025 study intentionally fed Whisper non-speech sounds to examine when the model generates text without speech input. The researchers observed recurring phrases such as “thank you” and “thanks for watching.” They suggested that these patterns may reflect common language in the web-sourced media used during model training. Because the experiments focused on non-speech files, their percentages should not be interpreted as expected hallucination rates for ordinary conversations or interviews.

Why AI Transcription Models Like Whisper Hallucinate

OpenAI describes Whisper as an encoder-decoder Transformer. Audio is converted into a log-Mel spectrogram and processed by an audio encoder, while a decoder generates the corresponding text tokens.

Because the model generates text sequentially, it can sometimes produce fluent output that is insufficiently grounded in the acoustic input. Researchers and developers have observed that this behavior is especially likely around:

Silence and long pauses: Extended non-vocal intervals may leave the decoder with weak acoustic evidence
Non-speech audio: Music, background sounds, and ambient noise can sometimes trigger text generation
Poor recording quality: Distortion, low volume, clipping, or strong background noise makes the speech signal harder to interpret
Speech disfluencies: Stuttering, hesitation, fragmented speech, and long pauses can increase recognition difficulty
Segmentation and decoding settings: Chunk length, thresholds, prompts, and other implementation choices can affect output

OpenAI reported that the original Whisper system was trained on 680,000 hours of multilingual and multitask supervised audio collected from the web. The broad dataset contributes to the model’s flexibility, but web-sourced training material may also expose the system to repeated captioning and video-language patterns.

Researchers have not established a single cause for every hallucination. The available evidence suggests that hallucinations emerge from an interaction among audio conditions, model architecture, training data, and decoding choices.

The Impact of Hallucinations on Transcription Accuracy

For professionals in healthcare, legal services, research, and media, a fabricated phrase may be more consequential than an ordinary spelling error.

Healthcare Consequences

In October 2024, the Associated Press reported that Nabla said its Whisper-based clinical documentation product was used by more than 30,000 clinicians across 40 health systems. Nabla also said the product had processed an estimated seven million medical visits.

The AP reported that Nabla erased original audio after processing for data-safety reasons. Nabla said clinicians were required to review and approve the generated notes. This example illustrates an important tension: deleting recordings may reduce some retention risks, but it can also prevent later comparison with the original conversation.

OpenAI has warned against relying on speech-recognition output in decision-making contexts where inaccuracies can produce significant consequences. Healthcare organizations should therefore require clinician review, establish clear retention and verification policies, and use services covered by appropriate privacy and contractual safeguards.

Legal and Research Implications

In legal work, fabricated transcript content could affect deposition preparation, case analysis, or evidentiary review. An automated draft should not be treated as an official record without the review and certification required by the applicable court, jurisdiction, or professional rules.

Research teams face a related risk. If a fabricated quotation enters qualitative coding or thematic analysis, it can influence findings and undermine the integrity of the study. Researchers should preserve source recordings, document their transcription process, and verify quotations used in reports or publications.

The Cornell-led study also found a disparity between its aphasia and control datasets. Hallucinations appeared in 1.7% of segments from speakers with aphasia and 1.2% of control segments. The researchers associated the difference with longer non-vocal durations. This does not establish that every person with a speech impairment will experience the same disparity, but it raises accessibility and fairness concerns for organizations using automated transcripts in consequential settings.

Best Practices for Minimizing AI Transcription Hallucinations

No mitigation method guarantees a perfect transcript. However, several practices can reduce risk and make errors easier to detect.

Technical Mitigation Approaches

Voice Activity Detection preprocessing: Removing or separating non-speech intervals before transcription can reduce the opportunity for a model to generate text during silence
Careful segmentation: Process audio in segments appropriate to the selected model rather than applying arbitrary duration rules
Conservative decoding settings: Review the documentation for the specific Whisper implementation and test settings against representative audio
Repetition and anomaly checks: Flag repeated phrases, abrupt topic changes, video-style sign-offs, or unusually fluent text during silent intervals
Version-controlled evaluation: Re-test workflows when changing models, language settings, prompts, preprocessing, or decoding libraries

A 2025 study also tested a post-processing method called a “bag of hallucinations,” which filtered recurring outputs produced in its non-speech experiments. The method reduced errors under those controlled conditions, but it should not be treated as a general substitute for source-audio review.

Workflow Best Practices

Maintain source audio: Keep recordings available until the transcript has been verified, subject to consent, privacy, retention, and legal requirements
Implement human review: Require review before high-stakes transcripts influence care, legal decisions, hiring, research findings, or publication
Prioritize high-risk passages: Review sections containing numbers, medications, names, dates, quotations, legal terms, or long pauses
Train teams on warning signs: Watch for repeated phrases, unexpected content, abrupt changes in style, or words appearing during silence
Document responsibility: Make clear who must review, approve, correct, and retain each transcript

Choosing Reliable AI Transcription Software

The Cornell-led research suggests that the hallucinations it identified were not reproduced uniformly across every speech-to-text service. When the researchers tested the 187 Whisper-hallucination samples with Google Cloud Speech-to-Text, Microsoft Azure Speech-to-Text, Amazon Transcribe, AssemblyAI, and Rev AI, those systems did not produce comparable hallucinated passages in that experiment.

This is a useful finding, but it is not proof that any service is incapable of hallucination. The comparison was limited to particular samples, service versions, and configurations.

When evaluating transcription software, consider:

Source-audio access: Can reviewers play the recording alongside the transcript?
Review efficiency: Are words linked to timestamps for rapid verification?
Speaker tools: Can users identify and correct speaker labels?
Terminology controls: Can teams add names, acronyms, and technical vocabulary?
Security documentation: Does the provider clearly document encryption, access controls, audits, data location, and retention?
Contractual suitability: Are appropriate agreements available for regulated or confidential data?
Export and audit options: Can teams preserve corrected versions and document approvals?
Representative testing: Has the service been tested on the organization’s actual languages, accents, microphones, and environments?

Price alone does not determine whether a system will hallucinate. Architecture, implementation, audio quality, review features, and operational controls all matter.

Leveraging Advanced Features for Enhanced Transcription Quality

Features surrounding the transcription model can make errors easier to identify and correct.

Editor and Review Tools

Word-level timecodes: Connect transcript text to the corresponding moment in the recording
Synchronized playback: Let reviewers listen while reading or editing
Speaker identification: Separate speakers and allow labels to be corrected
In-browser editing: Enable corrections without moving content into a separate application
Custom dictionaries: Let users add proper names, technical terminology, acronyms, and preferred spellings

Analysis and Insights

AI analysis tools can extract themes, summaries, chapters, topics, sentiment, or key moments from transcripts. These outputs remain dependent on the underlying transcript. Teams should correct important transcription errors before relying on downstream analysis, particularly when the results support research, reporting, or business decisions.

Collaboration Features

Shared workspaces, permissions, commenting, and review workflows can help teams assign responsibility and avoid version confusion. Collaboration features are most valuable when an organization also defines who is responsible for approving the final text.

Secure and Compliant AI Transcription for Sensitive Data

Security and compliance require more than a certification logo. Organizations should evaluate the service, selected plan, contractual terms, configuration, and intended use.

NHS England’s ambient-voice supplier-registry process requires participating suppliers to provide evidence of Class I medical-device status and a current Digital Technology Assessment Criteria assessment. This illustrates increasing scrutiny of clinical documentation tools, but it should not be interpreted as a universal rule applying identically to every transcription product in every NHS setting.

Security capabilities to examine include:

Independent controls audits, such as SOC 2 Type II
Encryption in transit and at rest
Role-based access controls
Multi-factor authentication and enterprise identity options
Data-location and retention information
Deletion and export controls
Business Associate Agreements or other appropriate contractual safeguards when required

Organizations should verify the exact scope of every certification and determine whether it applies to the service, plan, and use case they intend to deploy.

Why Sonix Supports Verifiable AI Transcription Workflows

For teams that need to review and manage transcripts efficiently, Sonix combines automated transcription with tools for checking the result against the source recording.

Current Sonix capabilities include:

A browser-based editor with synchronized playback, search, word-level navigation, and speaker labeling
Custom dictionaries that let users add names, technical terminology, acronyms, and preferred spellings
SOC 2 Type II certification, TLS encryption in transit, and AES-256 encryption at rest
HIPAA-oriented safeguards through Medical Sonix, including Business Associate Agreements for eligible healthcare organizations
Transcription support for 54+ languages
AI-powered analysis for summaries, chapters, themes, sentiment, and other transcript-based insights
Team collaboration controls, including workspaces, permissions, and sharing features

These controls help users detect and correct errors, but they do not eliminate the need for review. No automated transcript should be assumed to represent every spoken word perfectly without verification.

Whether you’re a research firm analyzing interviews, a newsroom working against deadlines, or an enterprise team managing large recording volumes, the safest workflow is one that keeps the transcript connected to the original audio and assigns clear responsibility for approval.

Final Verdict: Choosing the Right Transcription Solution for Your Needs

The decision between a self-managed Whisper workflow and a hosted transcription platform depends on technical resources, data requirements, review needs, and risk tolerance.

A self-managed Whisper workflow may be appropriate when:

The content is low risk and occasional errors are acceptable
The organization can test model versions, preprocessing, segmentation, and decoding settings
Technical staff can implement monitoring and mitigation
Source recordings remain available for verification
A qualified person reviews important output before use

A managed transcription platform may be preferable when you need:

A synchronized editor for comparing text with source audio
Speaker, timestamp, terminology, and correction tools
Shared workspaces and permissions
Documented security and privacy controls
Support for multiple languages and export formats
A workflow that lets nontechnical reviewers approve transcripts efficiently

For healthcare, legal, research, journalism, and other consequential environments, the key question is not whether a vendor promises perfect accuracy. It is whether the system makes errors visible, preserves the evidence needed to correct them, and supports a review process appropriate to the consequences of getting the transcript wrong.

Frequently Asked Questions

What is AI hallucination in transcription?

AI hallucination in transcription occurs when a speech-to-text system generates a word, phrase, or sentence that has no meaningful basis in the source recording. Unlike an ordinary substitution, a hallucination may introduce entirely new content. In one study of 13,140 short audio segments, researchers identified 187 hallucinations and classified 38% of them as containing explicit harms such as violent language, inaccurate associations, made-up information, or false authority.

Why does Whisper AI sometimes make up words or phrases?

Whisper is an encoder-decoder Transformer that generates text tokens conditioned on audio. When the acoustic evidence is weak—such as during silence, long pauses, background noise, music, or fragmented speech—the decoder can sometimes produce fluent text that is not sufficiently grounded in the recording. Training data, segmentation, prompts, model version, and decoding settings may also influence the result.

How can I reduce hallucinations in AI-generated transcripts?

Use Voice Activity Detection or other non-speech filtering where appropriate, test segmentation and decoding settings on representative recordings, flag suspicious repetitions or text appearing during silence, retain the original audio, and require human review for consequential content. Mitigation can reduce risk, but no preprocessing method guarantees an error-free transcript.

Are free AI transcription tools more prone to hallucinations?

Not necessarily. Hallucination risk is not determined by price alone. It depends on the model, architecture, implementation, audio conditions, and quality controls surrounding the system. Paid platforms may provide better review, security, collaboration, and audit features, but users should evaluate those capabilities directly rather than assuming that price guarantees accuracy.

Does Sonix address AI transcription hallucinations?

Sonix provides a browser-based editor synchronized with the source recording, word-level navigation, speaker tools, custom dictionaries, and collaborative review controls. These features help users identify and correct transcription errors. They should be treated as verification safeguards rather than proof that an automated transcript cannot hallucinate, particularly when the content will influence healthcare, legal, research, or other high-stakes decisions.

Get accurate transcription in minutes

Start transcribing smarter. Try Sonix free or explore our pricing to find the right plan for you.

Try Sonix Free See Pricing