To get a great sounding DAC its important to choose a DAC chip with lots of potential and ignore most of the datasheet specs. 'Having potential' means a simple as possible chip, without the usual bells and whistles which have been added to DAC chips in the past couple of decades. Going back to the 1990s most DAC chips were just that - DACs and had neither on-board digital nor analog filters using opamps. Both of these kinds of filters have the potential to screw up the sound so you do not want them on the chip where its impossible to bypass them.
Philips started a trend with their 'Bitstream' DACs of including on-chip filters and on-chip opamps (SAA7320). But for a while in the 80s and 90s they produced some excellent chips without all that extra fluff - for example TDA1541, TDA1543, TDA1545 and TDA1387. Burr Brown similarly had their PCM-series DACs, culminating in the PCM1704 where by most listeners' ears they'd already started to lose the plot, being seduced by numbers - PCM63 seems to be the pinnacle of their art. Analog Devices have had great designs too, the most recent being AD1865.
I've so far mentioned only audio-targeted DACs but there are others which are pure DAC chips - I've samples of an ancient one from ADI, the AD768. It's an example of a DAC targeted at the communications market, a field which in many ways holds more promise for great sounding D/A than does audio nowadays. TI/BB and Intersil also have offerings in this arena. Of the audio chips I've mentioned, all are out of current production so if you want to have a viable manufactured product you have to look beyond audio to find your chip. Schiit did exactly that in their multibit DAC designs, going to the medical/industrial segment. Metrum too started off with an industrial DAC chip from TI/BB but since have gone over to custom resistor arrays buried inside modules. Hobbyists though aren't constrained by production volumes and so have rarely had it so good with the wide choice of recycled devices on the secondary market.
I must confess I have a real soft spot for Philips (now NXP or even Nexperia) - partly because my first few CD players were Philips (they and Sony together being the innovators of the CD format) but also because their IC designs rock, and not just in DACs. Take in the realm of amps their in-car offerings (TDA8566 for example - I have a design for an amp using it here : Hi end chipamp). TDA1541 has been done to death by DIYers (Thorsten Loesch and Pedja Rogic have commercial product too) - how to choose from its many tweaks? TDA1545 has some DIY designs extant (Peufeu's 'Extremist' is here - TDA1545 DAC) - having listened to my own designs I've abandoned this chip on SQ grounds. TDA1543 has literally dozens of DIY projects and commercial boards/boxes (Lite DAC-AH is my favourite). Strangely TDA1387 has been largely ignored to date - there's a few units on Taobao : L1387 DAC but nothing in the audiophile mainstream. In regard to the tech behind the chips, the TDA1387 takes some beating as it continuously recalibrates itself on the fly so doesn't require the usual extreme resistor matching achieved in other manufacturer's offerings by laser trimming. Its also absurdly cheap as the chips are recycled from old 'Soundblaster' PC cards.
Next up - NOS is my choice here for lowest BOM cost. NOS is a TLA for 'No oversampling' and the craze of NOS began with a Japanese guy by the name of Kusunoki (go here for his web presence, I'm not thereby endorsing his technical justifications - http://www.sakurasystems.com/articles/Non-oversampling-DAC.html) who wanted to bypass the digital filter in a DAC. In the history of Philips' CD players, oversampling has always been used, initially because when CD came out they only had production ready a 14bit DAC (TDA1540) and CDs have 16bits worth on them. So oversampling (4X) was used as a way to gain a couple of bits from the DAC through averaging out 4 faster samples to create one slower one. When the TDA1541 was eventually introduced a couple of years later with 16bit performance, Philips could have deleted the digital filter (SAA7220) but to do so would necessitate a much more complex analog filter to keep out the inherent ultrasonic images produced by any DAC. Going NOS means the DAC's images start at just 24.1kHz (for 4XOS about 150kHz). So Philips stuck with the digital filter because the analog challenges were immense. Sony OTOH had a single 16bit DAC time-shared between two channels, I think they may have had a hybrid module taking care of anti-image filtering.
DIYers who've embraced NOS generally say no anti-image filtering is needed, some go further and say its detrimental to the sound. I've always found those claims too big to swallow so I did some experiments with high order passive (LC) filters, some of which I documented here : http://www.whatsbestforum.com/showthread.php?8858-Digital-that-sounds-like-analog. The short answer - for those who don't want to navigate that long thread - is - I found the filtered DAC to have not only a sweeter HF but, and this was most surprising, a much more 'holographic' presentation overall. By which I mean the acoustic space of the recording seems to become 3D. This latter feature is most subjectively alluring, once you've heard this kind of sound you really can't go back. I associate it with a reduced noisefloor electrically meaning that very low-level cues in the recording aren't getting garbled. Passive filters like the one I designed aren't at all commonly found in commercial DACs which to me is quite surprising. Nowadays inductors seem to be the Cinderella components, EEs typically think designing with them is some kind of black art. The only DAC I know of with a more complex filter than mine is the Zanden and this has a rather different design objective from my own. Needless to say its very highly regarded subjectively - if you follow that WBF thread you'll see one WBF member (Lloyd) has one.
Having determined to my satisfaction that fairly steep, aggressive filtering of DAC images was an essential component of a top sound quality DAC this rather sets me on a different path from almost all commercial and DIY DACs. Not being at all scared of being a lone explorer I looked into ways to get around the need for inductors. Active filters are a standard item in most DACs, generally they're implemented with opamps. But opamps don't sound particularly great and they sound even worse when fed ultrasonic frequencies, liberally available at the output of DACs. I have built a few DACs with active filters made out of discrete transistors, these sounded very good but the steepness factor I was able to achieve wasn't particularly high. A Q of about 10 was about as far as I was able to take my Sallen-Key based design in practical terms. Even this relatively low Q needs careful selection of capacitor values with an LCR meter. So if we have to select caps to get the filter response we need, why not inductors? Caps and resistors (the ingredients of active filters) are available in higher tolerances though, 1% is achievable but typically inductors can't be found better than 10% unless you wind them yourself.
When inductors are cheap and you're making a few DACs a solution is to buy a lot more than you need and to sort them out into bins and use all inductors from the same or adjacent bins, adjusting capacitors in combination to get a nice flat response. This is the way I've been building LC filters over the past few years - as I've not found a better solution that's how this DAC is getting built. The inductors which deliver the best bang for the buck are TDK's SLF7045 range. I've searched far and wide and these hit the sweetspot in many ways, not least the price in China is hugely lower than from Western distributors (Mouser for example has them over $1 whereas on Taobao about 10c). What makes these inductors stand out is they're relatively flat in their losses over the audio band. I've tried cheaper ones (Sumida is in general cheaper) but for some reason the one I tried had 4X higher losses at 20kHz than at 1kHz. Frequency dependent losses translate into high frequency droop and such losses are very hard to simulate in LTSpice (my simulator of choice).
There is a more 'hair-shirt' way to build high Q inductors and that's from ferrite beads. I have built and characterized a ferrite bead filter - the Q obtainable is really excellent, far higher than achievable with COTS parts. But each bead does not contribute much inductance, typically 1uH. So to build a 100uH inductor takes 100 beads in a string, rather impractical. One way to deal with this is to scale the working impedance of the filter. For an audio band low pass filter, a rule of thumb is that the working impedance (in ohms) is the inductance (in uH) divided by 10. So a filter with 100uH inductors has roughly a 10ohm working impedance - the working impedance needs providing at at least one end of the filter, sometimes both. If we were to re-design a 10ohm filter to work at 1ohm we scale all the capacitor values up by a factor of 10 and the inductors come down by the same factor. So a 1ohm filter can be built using 10uH inductors which are practical (just 10 beads each of which is 4mm long). More capacitance is required though which costs more money so I reckon there's a sweet spot impedance somewhere.
A few years ago I bought a bag of 4mm beads (1000 for 100rmb) from the local electronics market but I've since found (searching only yesterday) that on Taobao there are sellers going lower in price. The best bang for the buck is (so far) with 6mm long beads, I can buy 5000 (just over 1kg) for 100rmb.. I reckon that a 6mm bead is going to be a bit higher in inductance than a 4mm (have to await the delivery of my 1kg of beads to confirm this) - say its 1.5uH. Then we need at least 60mm length to achieve 15uH. Layout considerations mean its best to have the input and output of an inductor adjacent which means a single inductor ought to be built from an 'outward' and 'return' string of beads. Which gets us to 30uH, this should entail roughly a 3ohm load. This 30uH inductor comprises 20 beads which is 0.4mb. Roughly half the price of the TDKs. I am so far assuming that the cost of the wire running through the beads is negligible, I do need to verify this. (Cu rho is about 9, wire is 16AWG 1.3mm conductor diameter so roughly 100mm^3 of metal. Wire cost slightly under twice Cu spot price we've 0.9g per length, say $11/kg). That's about 0.1rmb for the wire, not insignificant here - the total solution is 0.5rmb vs the current price of 0.7rmb for the inductors. There is one advantage though that the value can be controlled better - with the TDKs we do need to buy more than we're going to use.
The consequences of working at a lower impedance are that the I/V stage (which sits directly at the output of the filter) needs to present a matched impedance to it, hence 3ohms. In turn this means it runs at a higher bias current, leading to shorter battery life. But what about the capacitors? (stay tuned)