DTMF Decoder

This decoder follows the classic DTMF detection architecture: audio is decoded to 8 kHz mono, then a bank of Goertzel filters runs over 205-sample frames — the block size that minimizes frequency-bin error for the 8 Q.23 tones — with extra second-harmonic filters to reject speech falsing. Every detected keypress reports its start time, duration, both measured frequencies (parabolic-interpolated to sub-bin precision), twist (high-group power relative to low group), frequency deviation in percent, and a confidence score. Acceptance rules mirror ITU-T Q.24 receiver guidance: tones must last ≥ 40 ms (adjustable), pauses ≥ 40 ms, frequency deviation within ±1.8% (strict mode: ±1.5%), deviations beyond ±3.5% rejected, and twist inside −8…+4 dB with receiver-style margin. The A–D fourth-column signals can be toggled off for real-world recordings where they only cause false hits. Honest scope: this decodes clean and moderately noisy DTMF (IVR prompts, voicemail, film props); it is not a telephony-grade talk-off-hardened decoder.