I've been noticing something in my call recordings that I'm hoping others might have insight on. The volume levels between the caller and the assistant are pretty dramatically different — the assistant comes through loud and clear, but the caller audio is consistently much quieter. I'm wondering if this imbalance might be affecting transcription accuracy, since the model is working with a weaker signal on the caller side. Has anyone else run into this? And if so, is there a way to boost the gain/levels on the caller channel specifically? Would love to hear how others have approached this. Thanks!

Hi there, Thank you for your message. Our team is currently out of the office. We operate Monday through Friday, from 9:00 AM to 8:00 PM Pacific Standard Time (PST). We’ll get back to you as soon as possible during our normal business hours. If your message is urgent, please mark it accordingly or include “URGENT” in the subject line, and we’ll do our best to respond promptly. Warm regards, Vapi Customer Support Team

Yeah, you’re not imagining it, this happens a lot. When the caller audio is much quieter than the assistant, it can absolutely hurt transcription because the model is working with a weaker signal. In VAPI, this usually comes down to how the inbound (caller) audio is gained or normalized before it hits transcription. The key is fixing it upstream — boosting or normalizing the caller channel before ASR, not after the call is recorded. I’ve helped teams clean this up with proper gain control and audio preprocessing, and it usually makes a noticeable difference pretty quickly. Out of curiosity, are these calls coming from PSTN (Twilio/Vonage) or web audio? That’ll point to the right fix. @Will

Thanks for the response! These calls are all going though Twilio.

That helps a lot, since everything is going through Twilio, this is a pretty common PSTN issue. With Twilio calls, the caller audio often comes in quieter because of how Twilio handles inbound gain, μ-law encoding, and automatic normalization before it reaches VAPI. The right way to handle it is to fix the input signal before transcription, not after the recording is saved. I usually audit the Twilio media stream, check the actual audio levels, and apply controlled gain or normalization so the ASR gets a much cleaner caller signal. That’s where transcription accuracy usually jumps. @Will

Hi Will, Thanks — this is a common Twilio/PSTN symptom. Quick guidance and actionable steps: 1. Root cause (most likely) • Twilio’s PSTN path uses μ-law + carrier-side gain/normalization; inbound caller leg often arrives quieter than agent TTS, hurting ASR. 2. Recommended fixes • Preprocess audio before ASR (don’t rely on post-call recordings): • Inspect Twilio Media Stream frames to confirm RMS/peak levels for the caller channel. • Apply controlled gain/normalization (e.g., +6–+12 dB or RMS normalization) to the caller channel prior to sending to the transcriber. • Use a limiter to avoid clipping after boosting. • Ensure you send the boosted PCM (mono, 16-bit LE, 16kHz) to Vapi/transcriber. • If using Twilio: consider switching transport/sample-rate settings or adding per-call gain in your Twilio Media Stream consumer. • If μ-law artifacts persist, enable Twilio dual-channel recording / debug recordings to view raw legs separately. 3. Short-term workarounds • Increase transcriber sensitivity/alternative model: try Deepgram Nova or Nova3 with keyterms if vocabulary is stable. • Use model-specific features (Deepgram Flux/Assembly end-of-turn) less aggressive while you tune levels. • Route test calls through a preprocessing step (your service) that normalizes caller audio then forwards to Vapi. 4. Diagnostics I can run if you share one recent Call ID (under 14 days) • Inspect Vapi/Twilio media stream logs and waveform levels for both legs • Confirm sample rates/codecs, per-call packet counts, and whether audio was received at low amplitude • Recommend exact gain multiplier/normalization target Reference: Voicemail/Call debugging and media guidance in our docs — see Debugging voice agents (Observe > Call Logs) for call IDs and recordings. https://docs.vapi.ai/debugging If you want, share one recent Call ID and I’ll pull the media diagnostics and propose exact gain settings.