BuyerVoice.AI

Phase 0 device tests

Seven short checks that decide three design questions before the platform build starts: does recording work on an iPhone, does Deepgram handle Safari audio and Spanish, and which moderator voice sounds right. Run this on an iPhone in Safari first, then on any other device you have. Copy the results log at the bottom and paste it back to Claude.

00Setup

Tests 04 and 05 call paid services, so they need the passcode. Tests 01 to 03 and 06 work without it.

Passcode set on server: checking Deepgram key: checking OpenAI key: checking

01Browser and recording support

Shows what this browser can record. Chrome and Android produce WebM, Safari produces MP4. The platform must accept both.

Loading...

02Audio unlock and playback

Mobile browsers only play sound after a tap. Tap the button. You should hear a short two-note tone. On an iPhone, try it once with the silent switch on and once with it off.

Not run yet.

03Record, pause, resume

Speak for about 10 seconds, tap Pause, wait a few seconds, tap Resume, speak again, then tap I'm done speaking. The playback should hold both halves as one continuous recording, and the live-time counter should not count the pause. Then try it again and switch to another app mid-answer without pausing, to see how the browser handles an interruption.

Tell me about the last time you chose a new piece of software for your team.

Not run yet.

04Transcribe the recording (Deepgram Nova-3)

Sends the recording from test 03 to Deepgram exactly as the browser produced it. Record once in English and once in Spanish. Check the words are right and note the round-trip time.

Record something in test 03 first.

05Moderator voice samples (OpenAI gpt-4o-mini-tts)

Generates the same question in each voice you tick, so you can compare them side by side and rate each one. For Spanish, a native Latin American listener should make the call.

Voices to compare

06End-of-turn detection (Silero VAD)

This is how the moderator knows you have finished speaking. Start it, answer the question below with natural pauses, then stop talking. It should not fire during a mid-thought pause, and it should fire shortly after you truly finish. Try 1.5 seconds first, then adjust.

What mattered most to you in that decision, and why?

idle
Not run yet. The first start downloads the speech model (about 2 MB).

07Largest text size

Tap A+ at the top of the page twice. Scroll the whole page. Every button must stay fully visible and tappable, nothing may overlap or run off the screen. Then tap the result below.

08Results log

Everything above writes here. Copy it and paste it into the Claude conversation.