Skip to Content
Noctune docs are in preview.
ReferenceAudio Quality

Audio Quality

Noctune’s transcription accuracy depends on the audio it receives. This page covers what makes a clean recording and how to avoid the most common problems. The underlying transcription engine (Deepgram) accepts a very wide range of inputs, but a few habits make a noticeable difference.

Quick checklist

  • Mic close, mouth forward. 6–12 inches from your mouth is a good target.
  • Record in a quiet space. HVAC, dishwashers, and hallway chatter degrade accuracy more than you’d expect.
  • One speaker at a time when possible. Simultaneous speech is the #1 cause of garbled transcripts.
  • Keep levels consistent. Avoid shouting then whispering — the AI normalizes, but extremes introduce errors.

Sample rate

Noctune handles any sample rate Deepgram does — commonly 8000, 16000, 24000, 32000, or 48000 Hz. In practice you don’t need to think about this; recording at your device’s native rate is correct. Don’t upsample telephony audio (8 kHz) to 16 kHz — it introduces artifacts that hurt accuracy.

SourceNative rateGood for Noctune?
iPhone / Android voice memo44.1–48 kHzYes
Browser getUserMedia48 kHzYes
Bluetooth headset (HFP)8–16 kHzYes, but avoid when possible
Desk phone handset8 kHzYes — don’t resample

Format and codec

Any of these work:

  • MP3, M4A (AAC), AAC — what most phone voice memos produce
  • WAV (PCM / Linear16), FLAC — higher bandwidth, no compression artifacts
  • Ogg, Opus, WebM — common for browser recording

If you have a choice, favor FLAC or WAV (lossless) for archival clarity and M4A/MP3 for smaller files. AAC and MP3 at reasonable bitrates (96 kbps+) are indistinguishable from lossless for transcription purposes.

Channels

Mono is fine and what most voice recorders produce. Stereo offers no accuracy benefit unless each speaker is on a separate channel, which is rare for in-room recordings.

Reducing noise

Things that noticeably help:

  1. Clip-on lavalier or headset mic over built-in laptop/phone mics — closer to mouth, less room.
  2. Carpet, curtains, soft furniture dampen reverb; tiled exam rooms are the worst case.
  3. Turn off ceiling fans, radios, and diagnostic equipment during the appointment when feasible.
  4. Don’t cover the mic. Pockets muffle the mid-range frequencies speech relies on.

Length

There is no lower bound on length. Upper bound is covered in Supported Formats & Limits.

Common symptoms and fixes

SymptomLikely causeFix
Words dropped or mergedMic too far / mouth turned awayMove mic closer; face the mic
Repeated words (“the the”)Echo / reverbRecord in softer room
Owner’s speech missingOne-sided mic placementPlace mic between you and the owner
Drug names wrongFast speech / background noiseSpeak drug names slightly slower; add to templates
Breed names wrongUncommon breed + noiseSpell out once at start of visit

Sources: Deepgram Supported Audio Formats , Determining Your Audio Format .

Last updated on