AI transcription; local vs cloud

AI transcription turns spoken words into written text using automatic speech recognition, or ASR.

The recording is divided into small sections and converted into a numerical representation of the sounds. An AI model then identifies patterns and predicts the most likely words, punctuation and timings. A separate process can help distinguish between speakers.

The model is not listening like a person. Background noise, overlapping speech and unfamiliar terminology can still cause errors, so transcripts and generated notes should always be reviewed.

Local and cloud transcription

Local and cloud transcription use similar AI technology. The difference is where the model runs and where the recording travels.

With cloud transcription, audio is uploaded to a provider’s servers. Those servers create the transcript and return the text. This works across many devices and gives providers access to substantial computing power, but the raw recording must leave the therapist’s computer.

With local transcription, the model runs directly on the device. The recording can become a transcript without being uploaded to a transcription provider.

AirNote uses local transcription. Session audio is processed on the therapist’s Mac and is not sent to AirNote, Apple, OpenAI or another cloud provider for transcription. It is deleted after transcription completes successfully.

Why this matters for therapy PHI

A therapy recording can contain a client’s voice, health history, symptoms, relationships and the full context of a private conversation, as well as the therapist's voice and contribution to the conversation.

For US therapists covered by HIPAA, identifiable information about a client’s mental health or care may be protected health information, or PHI. This protection can apply to oral as well as written information, as explained in HHS guidance.

Cloud processing can be secure and compliant. Suitable providers can use encryption, access controls and Business Associate Agreements. However, compliant processing still means another organisation receives the recording.

Every additional system introduces another transfer, processing environment and set of security controls. Local transcription removes that transfer entirely.

This is data minimisation in practice: do not send the most sensitive material elsewhere when the work can be completed safely on the therapist’s device.

Apple silicon changed what is possible

High-quality AI once required specialist servers. Apple silicon—the M-series chips in present-day Macs—has changed that.

These chips include a Neural Engine designed for machine-learning work. Apple’s technology can run models across the Mac’s processor, graphics hardware and Neural Engine.

Transcription models are also smaller and more focused than general-purpose AI systems. They perform one specialised task: converting speech into text. This makes it practical to run them quickly on a Mac.

Local does not mean less accurate

Where a model runs does not determine its quality.

Accuracy depends on the model, its training data and how the transcription system handles long recordings, silence, speaker changes and overlapping speech.

Modern local models can provide accuracy on par with—and often better than—cloud transcription models.

This does not mean every local model beats every cloud service. It means high-quality transcription no longer requires audio to leave the device.

Local processing keeps costs lower

Cloud transcription creates a cost for every minute of audio. It's one of the more expensive forms of AI processing. Consequently, providers must pay for server processing, data transfer and the infrastructure behind each transcription.

Local transcription uses computing power the therapist already owns. Once the model is installed, AirNote does not need to purchase cloud transcription for every session.

AirNote still pays for secure cloud services used by its separate text-based AI features and other product operations. Removing cloud audio transcription, however, eliminates a significant recurring cost.

That helps AirNote remain ultra-affordable. Comparable paid plans from large AI scribe providers are around four to six times AirNote’s price.

A better default for therapy audio

Cloud transcription can be appropriate when suitable local hardware is unavailable or an organisation needs centrally managed systems.

But therapists using a modern Mac no longer need to upload an entire therapy recording simply to turn speech into text.

Local transcription keeps sensitive source material closer to the therapist, reduces unnecessary data transfers and lowers the cost of providing the service.

For therapy, it is a more private and more affordable place to begin the documentation process.