Abstract. We present a spectroscopic analysis of a sample of 21 galaxies with z = 0.2−1.5 drawn from a 25 square arcmin ultradeep ISOCAM survey at λ eff = 15 µm centered in the WFPC-2 Hubble Deep Field South. Near-infrared spectra are reported for 18 ISO sources, carried out with ISAAC on the VLT, aimed at detecting the redshifted H α + [N]. Additional optical data come from the ESO VLT/FORS2 and NTT/EMMI, primarily targeting [O], [O] and H β for further physical insight. Although not numerous in terms of areal density in the sky, this population of very luminous IR sources has been recently found to be responsible for a substantial fraction of the extragalactic background light energy density. H α line emission is detected in virtually all the observed objects down to a flux limit of 7 × 10 −17 erg cm −2 s −1 (corresponding to L Hα > 10 41 erg s −1 at z = 0.6 for H 0 = 65, Ω Λ = 0.7 and Ω m = 0.3). Our analysis (including emission line, morphology, and SED properties) shows clear evidence for AGN activity in only two of these sources: one type-I (with broadened H α at z = 1.57) and one type-II quasars (with inverted [N]/H α ratio at z = 1.39), while we suspect the presence of an AGN in two further sources (an Ultra-Luminous IR Galaxy, ULIRG, at z = 1.27 and a luminous galaxy at z = 0.69). The H α luminosities indicate star formation rates (S FR) in the remaining sources between 0.5 and 20 M /yr, assuming a Salpeter IMF between 0.1 and 100 M and without extinction corrections. We find good correlations between the mid-IR, the radio and H α luminosities, confirming the mid-IR light as a good tracer of star formation (while the S FR based on H α flux show some large scatter and offset, which are still to be understood). We have estimated the baryonic masses in stars with a newly-developed tool fitting the overall optical-IR continuum, and found that the host galaxies of ISO sources are massive members of groups with typically high rates of SF (S FR ∼ 10 to 300 M /yr). We have finally compared this ongoing SF activity with the already formed stellar masses to estimate the timescales t SF for the stellar build-up, which turn-out to be widely spread in these objects between 0.1 Gyrs to more than 10 Gyr. The faint ISOCAM galaxies appear to form a composite population, including moderately active but very massive spiral-like galaxies, and very luminous ongoing starbursts, in a continuous sequence. From the observed t SF and assuming typical starburst timescales, we infer that, with few exceptions, only a fraction of the galactic stars can be formed in any single starburst event, while several of such episodes during a protracted SF history are required for the whole galactic build-up.