Recording specs

Should speech data be recorded at 16 kHz or 48 kHz?

Updated 2026-08-01 · 4 min read

Short answer

Record at the highest rate your deployment could ever need and downsample later — in practice 48 kHz / 24-bit for studio work, because you can always produce 16 kHz from 48 kHz but never the reverse. The exception is telephony: if your model serves an 8 kHz narrowband contact-centre channel, capture over a real telephony path rather than downsampling studio audio, because band-limiting a clean recording does not reproduce codec artefacts, packet loss or handset response.

Key takeaways

Capture high, deliver at the rate your pipeline wants — downsampling is lossless in practical terms, upsampling is not.
Telephony data must be captured over a real telephony path, not simulated by downsampling.
Bit depth matters as much as sample rate for headroom during post-processing.

Why 48 kHz for studio capture

48 kHz / 24-bit is the standard for TTS and for any corpus that may be reused. It preserves the high-frequency detail vocoders use and gives enough headroom that a slightly hot take does not clip.

The storage cost difference is negligible against the cost of re-recording a corpus because a future model needed wideband audio.

Why 16 kHz is often enough for ASR

Most ASR front-ends operate at 16 kHz, and speech energy relevant to recognition sits well below 8 kHz. Delivering 16 kHz files keeps ingest simple and training fast.

Deliver 16 kHz derived from a 48 kHz master rather than recording at 16 kHz, so the master remains useful when the next model wants more bandwidth.

Annotators writing prompts and responses for LLM training data — recording specs context for Should speech data be recorded at 16 kHz or 48 kHz — Annotators writing prompts and responses for LLM training data

The telephony exception

An 8 kHz narrowband corpus for a contact-centre model must be captured through a real telephony path. Codec compression, jitter, packet loss and handset frequency response all leave signatures that a downsampled studio file does not have, and models trained on the simulated version underperform on the real channel.

Where a project needs both, we record parallel studio and telephony captures of the same speakers so the pair can be used for channel-robustness training.

What to write in the specification

State the capture rate, delivery rate, bit depth, channel count and whether telephony audio is real or simulated. Ambiguity here is a frequent cause of a corpus being rejected at acceptance for reasons that had nothing to do with the speech itself.

Frequently asked questions