If you look at a phoneme set for English (e.g. Wikipedia's IPA for English), that lists the phonemes found in English, excluding dialectal variations. As such, you need a set of sentences that include at least one word listed in the given phonemes.
The dialect being used is important, because that affects the distribution of phonemes. This includes splits, mergers and sound shifts.
John Wells has a reasonable analysis for vowels using Lexical Sets -- groups of words whose vowel share the same sound. Different lexical sets may share the same sound, in which case they are merged. For example, modern (non-conservative) English accents don't make a distinction between NORTH and FORCE. Also, American and British English use a different sound for the vowel in SQUARE and several other vowels; most of this is due to a rhotic vs non-rhotic distinction.
Thus, you need to take this analysis into account if you want to properly account for different accents. You will also need to extend the lexical sets if you want to account for Welsh, Scottish, Irish and Liverpudlian accents that preserve various Middle and Early Modern English distinctions, having resisted various mergers such as MEAT-MEET=FLEECE (
/iː/) and FIR-FUR-FERN=NURSE (
/ɜː/). An understanding of the historic splits and mergers is useful for understanding these accents.
Likewise, you need to account for other splits in progress, such as the BAD-LAD split found in Australian English and American and Canadian English accents with Canadian raising.
Wikipedia has a fairly comprehensive description of various accents and the different splits and mergers.
You also need to be aware that different people transcribe different sounds differently depending on what information they choose to preserve. You need to be aware of this if you choose different sources for phonetic information. For example, Wikipedia is not consistent across different articles. Likewise, the CMU pronunciation dictionary is not consistent -- it mixes different American accents and in some places mixes up phonemes.
A phoneme can have different realizations depending on the context and accent. For example,
/p/ is usually aspirated at the start of a word (
[pʰ]). The International Phonetic Alphabet chart for English dialects Wikipedia page has information on possible phones for given consonant phonemes.
Some of the phone realizations are the result of forward or backward assimilation, or other processes. For example
/m/ is typically realized as
To account for these, you should be interested in diphone coverage. If you want to modify phonemes based on their environment to model processes like assimilation, you should be interested in triphone coverage. It can work with diphone coverage, but it makes the processing more complex. Both of these increase the amount of units involved (although not all combinations are present in English).
For diphone coverage, you can look at diphone synthesizers. For example, the CMU arctic databases use 1150 utterances (1132 sentences) from several Project Gutenberg texts. These voices (including the Canadian, Scottish and Indian voices) have the phonemes transcribed in Arpabet, which is a US English phoneme set. Thus, the information in these voices don't preserve accent features (that is, the Scottish accent is Americanized, losing at least the FIR-FUR-FERN (
/ɛɾ/), LOCK-LOCH (
/x/) and WINE-WHINE (
/ʍ/) distinctions). I don't know how good the coverage is for non-General American accents.
The Rainbow Passage is used by the International Dialects of English Archive, so should have good phoneme coverage across accents.
The Speech Accent Archive uses an elicitation paragraph for comparing its accents. This paragraph ("Please Call Stella") is found in footnote 1 on the about page.
Another elicitation paragraph I have found was The Tiger and the Girl.
You can search google for other elicitation paragraphs.