A type of vocoder is described that promises modest bandsaving and elimination of the pitch-tracking and voiced-unvoiced switching inherent in spectrum channel vocoders. A speech signal f(t) suffers little degradation when passed through a parallel bank of contiguous bandpass filters and then recombined. If fn(t) is the output of the nth bandpass filter, the original signal f(t) is approximated by Σnfn(t). Each fn(t) can be represented by two parameters: the value of the short-time amplitude spectrum of f(t) evaluated at frequency ωn, and the time derivative of the short-time phase spectrum also evaluated at ωn. These data are transmitted to a synthesizer that produces approximations to each fn(t). The complex short-time spectrum calculated for each channel is F(ωn,t) = ∫ −∞∞ ∫ (λ)h(t−λ)e−jωnλdλ=|F(ωn,t)|eiφ(ωn,t), where h(t) is the impulse response of a realizable low-pass filter. The amplitude |F(ωn,t)| and the phase derivative φ̇(ωn,t) are formed for each channel, low-pass-filtered and transmitted to the sythesizer. At the synthesizer, φ̇(ωn,t) frequency-modulates an oscillator of nominal center frequency ωn, and |F(ωn,t)| amplitude-modulates the same oscillator. The synthesized signals for all n channels are then summed. In effect, the |F(ωn,t)| signals carry the spectral envelope information, and the φ̇(ωn,t) signals carry the excitation information. Previous experience with channel vocoders shows that the |F(ωn,t)| signals may be band-limited to around 20 cps. Our experiments with the phase vocoder indicate that the φ̇(ωn,t) signals may be similarly band-limited. Speech transmitted by the phase vocoder is demonstrated.