Testing Feature Interaction: Between-stream Irrelevant Speech Effects in Immediate Recall

Immediate serial recall of visually presented lists is disrupted by irrelevant background speech. One explanation for the irrelevant speech effect assumes that features of the auditory material become incorporated into the memory trace of the to-be-remembered item thereby reducing the fidelity of the short-term trace. From this perspective the resultant short-term memory trace is a composite of features of the list item and features of an item in the irrelevant stream. While there is evidence that item interactions in short-term memory are observable, there is currently no direct empirical evidence for such interactions involving irrelevant speech. We report six experiments using a short-term cued recall task that manipulates proactive interference in which item interactions have been observed. In these experiments we consistently show that with irrelevant speech specific items in the auditory stream influence target recall and the presence or absence of proactive interference. These results provide relatively direct evidence for feature interaction involving irrelevant speech. The results are evaluated against the three current models of irrelevant speech effects Feature Interactions 3 Testing Feature Interaction: Between-stream Irrelevant Speech Effects in Immediate Recall Ordered recall of short lists of items is disrupted by someone speaking in the background, even though the instructions stress that this speech is to be ignored. The irrelevant speech effect, as this phenomenon is known, has become an important benchmark in the development of theories of short-term recall. Currently, there are three different explanations for the irrelevant speech effect. The working memory explanation (Baddeley, 1986) argues that the words in the unattended stream are registered in the same short-term store as the list items and this produces recall problems. At this point in time Baddeley has not specified what causes the difficulties. The initial versions suggested that irrelevant speech corrupted the phonological store in some nonspecific way but this view was rejected in favour of an assertion that irrelevant items add some form noise to the recall process (Baddeley & Hitch, 1994). In the most recently published discussions of the phonological loop model, Baddeley still gives no clear indication of precisely how irrelevant speech produces a recall deficit (Baddeley, 2000). The O-OER model (Jones, 1993) argues that the interfering effects of background sound are not limited to speech. Jones and his colleagues have consistently shown that tones produce the same pattern of disruption as speech. In both instances disruption is a function of the degree to which items in the irrelevant stream change from item to item; changing state stimuli produce more interference than steady state stimuli. In the O-OER model, the deficit from irrelevant sound is due to confusions in pathways that maintain the order of items in memory. When a list is learned the items are linked to each other to form a pathway from the first to the last item. In the case of irrelevant speech, the items in the auditory stream form a second pathway that links the irrelevant items together. At recall the participant attempts to recover the list items by following the pathway for those items. However, confusions in the Feature Interactions 4 linkages are possible which produce recall errors. Such cross talk between the two pathways is more likely with changing state sound than steady state sound. In the Jones model the memory trace for the list items is not degraded in any way by the irrelevant speech. In the Baddeley model it is not clear how or if trace degradation occurs. In the third account, irrelevant speech directly affects the memory trace. Neath's (1999; 2000) adaptation of the feature model (Nairne, 1990) argues that features from the irrelevant speech become absorbed into the memory trace of the list items. That is, features of the words in the auditory stream interact with features of the list items such that the memory trace for the list item is changed. The resultant short-term trace loses its fidelity and exists as a corrupted version of the original trace. The corrupted short-term trace is then matched to a series of long-term traces to select an appropriate response. To the extent that the short-term trace is corrupted, the probability of selecting the correct item decreases in that there may not be a substantial match with any list item, or the match might be closer to an incorrect item than to the correct item. In discussing Neath's approach, both Baddeley (2000) and Jones and Tremblay (2000) implicitly argue that there is no direct evidence for feature adoption. Furthermore, they suggest that at face value, feature adoption predicts that the similarity between the auditory items and the to-be-remembered items should have an impact upon performance. Both expect that the greater the level of similarity, the greater the resulting disruption. The available empirical evidence indicates that between-stream similarity has little impact upon recall. There are now close to a dozen separate experiments that have explore between-stream similarity effects (Bridges & Jones, 1996; Jones & Macken, 1995; Larsen, Baddeley & Andrade, 2000; LeCompte & Shaibe, 1997) and none, bar Salamé & Baddeley's (1982) Experiment 5, have found any evidence in favour of between-stream similarity. The implication is that this lack of evidence is problematic for the feature model. Although Neath Feature Interactions 5 is able to formally model the absence of between-stream similarity effects, neither Baddeley nor Jones and Tremblay are overly receptive to the assumptions Neath has made in order to do so. As such, the absence of between-stream similarity effects has the potential to be problematic for feature adoption. This issue might be partially resolved if between-stream similarity effects could be observed. In the current research we test the ideas behind the Neath perspective, rather than the way feature adoption is formally implemented in the feature model. That is, the primary interest is in the possibility that items in memory may interact with each other and that a resultant memory trace may in fact be a blend of two or more items. We think that Baddeley and Jones and Tremblay are incorrect in assuming that feature adoption can only result in a detrimental effect upon performance. It may do so, but it is possible that it may have no effect or it is possible that irrelevant speech may facilitate recall under some circumstances. Some examples might make this clear. The feature model involves a matching process in which an easily corruptible shortterm trace is matched to a set of traces in long-term memory where both short-term and longterm traces consist of sets of features. The item that produces the best match is selected for recall. Let us take the simplest case where two words have been studied, say dog and cat, and the person is matching the short-term trace of cat to the long-term traces. If one were to do this task solely on the basis of the match between the letters of the respective words, it is readily apparent that there is complete overlap between the letters involved in the short-term trace (c, a and t) and the long term trace for cat (c, a and t) and no overlap with the long term trace for dog (d, o and g). Consequently, there should be little problem in recalling cat. In the feature model forgetting is based upon retroactive interference where features in the short-term trace are overwritten. Now let us assume that the short-term trace has been overwritten such that only the middle letter is available (a). It is still the case that matching a Feature Interactions 6 to the features of c, a and t and d, o and g will still produce a better match for cat than dog so cat is again likely to be recalled. Now assume that with unattended speech features of the auditory words are absorbed into the short-term trace. The simplest form that this could take is if some of the features in the irrelevant words replace the features of the short-term trace in their respective positions within a word (Neath does not make this assumption). Consider the effects of having fib, cot or dig in the auditory stream, a corrupted short-term trace in which only a is available, and the two consonants of the auditory items are incorporated into the short-term trace in their respective positions. In the case where fib is paired with the short-term trace, the resulting short-term trace will be fab. This trace will match with the long-term representation of cat on one of the three features and will not match any features of dog. Cat is likely to be recalled. In the case where cot is paired with the short-term trace, the c and the t will be absorbed such that the resultant trace is cat. This will produce a perfect match with the long-term trace of cat. Here we have an instance of where irrelevant speech restores the missing features and actually facilitates recall. In the case of dig in the irrelevant stream, the d and the g will be absorbed into short-term trace to form dag. Note now that the short-term trace produces two matches with dog and only one with cat; dog is more likely to be recalled. In this instance the irrelevant speech is likely to produce a recall error. While the above examples maybe overly contrived, the general point is still valid. As Baddeley and Jones and Tremblay note, if items are represented in terms of features and features from different items are able to interact, then short-term recall should be sensitive to the type of features that are present in both memory and the irrelevant stream. However, it is not necessarily the case that the effects should be disruptive. It should depend upon the features involved. Sometimes list recall might be enhanced and in other instances it might be Feature Interactions 7 hurt by irrelevant speech. We now turn to evidence for the type of feature interaction effects that we are concerned with. Feature (Item) Interaction Effects Over recent years we have been exploring proactive interference (PI) effects in a shortterm cued recall task (Tehan & Humphreys, 1995, 1996, 1998; Tolan & Tehan, 1999). In this task participants are presented a series of trials consisting of either one or two four-word blocks. Participants are instructed to pay attention to all the items, but once they find out that a trial is two-block trial, they are directed to forget the first block and concentrate on remembering the second block because it will be this block that will be tested. At test a category cue is presented and participants are asked to recall the item from the most recent block that was an instance of the category. The critical trials in the experiment are the twoblock trials because it is on these trials that PI is manipulated. On the two-block trials there is always a target item presented amongst three unrelated filler items in the second block. In the control conditions, the target item is the only instance of the category in the list. In the interference condition, a second instance of the category is presented in the first block. Thus, the cue might be ANIMAL, the block-2 target might be cat, and the block-1 foil might be dog. Using this paradigm we have consistently observed PI effects on a delayed test (Tehan & Humphreys, 1995, 1996; Tolan & Tehan, 1999). Our exploration of item interaction effects has centred on immediate recall where sometimes recall is immune to the effects of PI and in some instances PI can readily be observed. It turns out that the presence or absence of PI critically depended upon the phonological similarity between the foil and the target (Tehan & Humphreys, 1995, 1998). Recall was immune to PI when taxonomic categories were used and the foil and target did not rhyme. However, when word ending cues were used (_AT as the cue, cat as the target, and hat as the foil) or when rhyming instances of taxonomic categories Feature Interactions 8 were used (ANIMAL, cat as the target, and rat as the foil) PI effects were observed. Target recall was depressed and recall of the foil increased substantially. Tehan and Humphreys (1998) tested for feature interaction by exploring the possibility that with non-rhyming instances of a taxonomic category, the features of the foil could be provided by another word. In these experiments they utilised standard control and interference trials but added a series of trials in which one of the filler items in the second block was a rhyme of the foil (ANIMAL, cat as the target, and dog as the foil, and log as a block-2 filler). When the rhyme was absent from the list, no effects of PI were observed. However, when the rhyme was in the list, PI effects were observed. The presence of log in the second block enhanced the recall of dog. They took the idea of feature interaction a step further by taking the phonemes of the foil and distributing these items across three filler items in the second block. Thus dog was decomposed into dart, mop and fig. This manipulation again produced an increase in PI, primarily through enhancing recall of the foil. The strengthening of the foil by items elsewhere in the list that share features with it is what we mean by an item interaction. The above research suggests that at the point of recall the features representing the target, the foil and the filler items appear to be simultaneously active in memory and the phonemic features of these items appear to interact with each other. We think that results of this type provide reasonably direct evidence for feature interaction. If features from filler items in a list can influence recall it is not a giant leap to suppose that irrelevant auditory items could have similar effects. That is, if dart, mop and fig were items included in the auditory stream and the d, o and g were absorbed into the memory trace, recall of the foil dog should be enhanced in the same way that it was when these words were filler items. In short, feature interaction assumptions predict that phonological similarity between the items in auditory and visual streams could enhance PI in the cued recall task in the same way and for the same reasons that irrelevant filler items in the list enhanced PI. As Feature Interactions 9 we indicated previously, the available empirical evidence indicates that between-stream similarity effects are not observable, at least as far as immediate serial recall goes. The failure to replicate the between-stream similarity effect appears to present a major obstacle for feature interaction assumptions. It looks as though similarity of features has no effect whatsoever. However, to this point, irrelevant speech effects have not been studied using a cued recall task, nor has the impact of irrelevant speech on short-term PI effects been studied. The experiments we report here first look at the effects of irrelevant speech on the cued recall task, and then explore the effects of between-stream similarity on the cued recall task. To preview our results we show that between-stream similarity does have an impact upon the cued recall task in that PI effects are influenced by the irrelevant items in the auditory stream. Experiment 1 This experiment explores the effects of irrelevant speech in a short-term cued recall task. In this experiment we wish to determine whether or not irrelevant auditory input has a detrimental impact upon cued recall. Irrelevant speech effects are readily observed in immediate serial recall where multiple responses are required and order must be maintained. In fact, Jones (1993) has argued strongly that irrelevant speech effects are limited to those tasks that involve maintenance of order. Since, there is no a priori reason for students to maintain order in the cued recall task, irrelevant speech effects may well differ to those in serial recall. Furthermore, irrelevant speech effects are less robust in tasks where a single response is required (Beaman & Jones, 1997). Provided that irrelevant speech does have an impact the question remains as to whether or not changing-state speech produces more of a decrement than steady-state speech (Jones & Macken, 1995). Thus, in the current experiment we compare a quiet control to both steady-state and changing-state speech. Feature Interactions 10 Finally, the disruptive effect of irrelevant speech seems to occur at storage or retrieval rather than at encoding (Baddeley & Salamé, 1986; Colle & Welsh, 1976; Colle, 1980; Miles, Jones & Madden, 1991). In the current experiment we present the irrelevant speech simultaneously with the list items and test immediately, or we present the visual items in the absence of speech but delay testing until two seconds of irrelevant speech has been presented after the final item in the list. Given prior findings, irrelevant speech effects should be equivalent in the two conditions. Method Participants Forty introductory level psychology students participated in this experiment for course credit. The first twenty participants were assigned to the irrelevant speech during input condition while the remaining twenty participants performed under speech during rehearsal period delayed condition. Materials The first step in constructing the stimulus set was the creation of two mutually exclusive pools of words to serve as materials for the critical two-block trials in the experiment. One pool contained the items that would serve as targets on the critical cued recall trials, while the other acted as a source for filler items. Following the Tehan and Humphreys (1995; 1996, 1998) procedure, the target items were created by selecting one instance from thirty-six different taxonomic categories was from the South Florida Category Norms (McEvoy & Nelson, 1982). These items were low dominant items within the category’s hierarchy. The filler word-pool consisted of 252 words that came from different taxonomic categories to those used in the selection of the critical trials. The remaining categories from the South Florida Category Norms (McEvoy & Nelson, 1982) and the Shapiro and Palermo (1970) category norms served as the sources of these items, with multiple items Feature Interactions 11 from each category being selected. This ensured that there was no overlap in category membership between critical and filler items, but it was possible that two items from a category could appear as filler items on a trial. Items for each trial were chosen without replacement from their respective stimulus pool. Each subject studied a unique set of forty-eight trials. Each trial consisted or one or two four-word blocks with the 36 two-block trials being the critical trials in the experiment. To create each two-block trial, seven items were randomly selected without replacement from the filler pool. Four of these items were randomly assigned to the first block and the remaining three were assigned to the second block. A target item was then randomly sampled without replacement from the target pool. On half the trials this item was placed as the second item in the second block and on the other half it appeared in position three. The thirty-six twoblock trials where then randomly assigned to the three irrelevant speech conditions such that equal numbers of trials were assigned to each condition. The study also contained 12 one-block trials. The items for these trials were also selected from the Sth Florida norms but were selected such that there was no overlap in category membership with any of the items in the two-block trials. The target items in these lists tended to be placed in the first and last serial position and were always tested in the absence of irrelevant speech. The order of the 48 trials was randomised for each subject. We generated the items for the irrelevant stream by creating 72 phonotactically legal non-words and then assigning three non-words to each of the 24 lists that were to be studied under irrelevant speech conditions. These non-words were either one or two syllables long. Each three non-word combination in the changing-state condition was phonemically dissimilar from each other (e.g., NUG-PROG-BULA) while the three non-word combination in steady-state condition was phonemically similar to each other (e.g., GEN-VEN-NEN). None of the non-words rhymed with any of the target items. Feature Interactions 12 It should be stressed that our manipulation of steady-state speech is not truly steadystate; the first phoneme changes across words. True steady-state speech would involve a simple repetition of a single item. We have adopted the current type of speech because this was the version of low-level changing-state speech that Jones and Macken (1995) used when exploring between-stream similarity effects. Procedure At the beginning of the experiment subjects were informed that they would be studying a series of one-block and two-block trials in which a block consisted of four words. However, it was also stressed that at any one point in time they only had to remember that most recent block of four items. Consequently, if the trial was a two-block trial, and this was signified by the presence of an exclamation mark (!) as the block separator, they were to forget the first block and concentrate on remembering the second block because it would be on this block that they would be tested. They were told that one-block and two-block trials would be randomly interspersed throughout the experiment and that since they would not know in advance what type of trial it was, it was in their best interests to treat each trial as a one-block trial until they learned otherwise. They were also told that on some trials they would hear the experimenter repeating a series of words and that they were to do their best to ignore this material. The events that the subjects were concerned with all happened in the bottom left hand corner of a computer monitor. Each trial began with a READY sign displayed for two seconds. The study items were then displayed individually in lower case at a rate of one word per second, and subjects were instructed to remain silent and not to move their lips throughout the presentation of the study items. On two-block trials, the block separator, (!), was presented for one second after the fourth word in the first block and before the first word in the second block. At recall, a category cue was presented in upper case for two seconds. On an immediate Feature Interactions 13 test the cue appeared immediately after the fourth item in the block. On trials in which a rehearsal period was employed, the screen went blank for two seconds after the final word had been presented and subjects were instructed to rehearse the last block of items. After 2 seconds the cue appeared in the same corner of the screen. With the appearance of the cue, participants were requested to verbally recall the item from the most recent block that was an instance of the category. Subjects had five seconds to make a response before the next trial began. The experimenter recorded the subject’s responses (correct recall, intrusion errors, omissions, etc) on a hard copy of the subject’s input file. The experimenter provided the irrelevant speech in the experiment. On the trials where irrelevant speech was manipulated, the items were presented in the top right hand corner of the screen and the experimenter repeated the items for the time they remained on the screen. With items appearing simultaneously in the diagonal corners of the screen, there is potential for visual activity to be a source of distraction and a possible confound. However, with the seating arrangements used, the perceptual experience for both the subject and the experimenter was at most a flicker in the peripheral visual field. It was not possible to read the words, nor was it possible to identify that the visual material was in fact words. Participants reported that this irrelevant visual material did not distract them and many in fact did not notice it. They mentioned that they were too busy concentrating on the task at hand and ignoring what was being articulated in the background. Consequently, any effects of the nature of the words in the irrelevant stream are unlikely to be due to the visual events that were occurring in the top part of the screen. In the case of irrelevant speech that was presented during input, the three non-words appeared in the top right hand corner as the first to-be-remembered word appeared in the bottom left hand corner and stayed on the screen until the final list word disappeared. During this time the experimenter repeated the three non-words at a rate that produced about eight or Feature Interactions 14 nine repetitions of the non-word sequence during the 9 seconds it took to present the study list. When the irrelevant speech was presented during the two-second rehearsal period, the three non-words appeared in the top corner immediately after the final to-be-remembered disappeared from the bottom corner. The non-words were present for two seconds and the experimenter pronounced the sequence only once. Thus, the amount of irrelevant speech differed in the two conditions. The non-word sequence was repeated eight or nine times in the immediate test condition but only once in the rehearsal period condition. Results and Discussion One-Block Trials The aim of these trials was to ensure that participants attend to the first block of words on all trials. The participants correctly recalled 86% of block-one trials in this experiment and performed at similar levels in the remaining experiments. As such these trials appeared to achieve their purpose and are not reported in subsequent experiments. Two-Block Trials Figure 1 displays the mean proportion of correct recall and errors made while attempting target recall under quiet, changing-state and steady-state conditions during input and rehearsal phases. However, in the interests of space analyses are only conducted on target recall. In places during the remainder of the manuscript we will make passing reference to the number of omission errors made. Where the text indicates that there are differences in omissions in the various conditions, these differences are statistically reliable. An alpha level of .05 has been used in all statistical analyses. Correct recall A 2 x 3 repeated measures analysis of variance was conducted on the target recall. A main effect for when irrelevant speech was presented was evident, F (1, 38) = 15.09, MSe = 0.05. A significant main effect for irrelevant speech also emerged, F (1, 38) = 61.06, MSe = .01 Planned comparisons indicated that correct recall was significantly superior Feature Interactions 15 under the quiet condition than the steady-state, t (39) = 4.77, and the changing-state conditions, t (39) = 6.77. However, the changing-state and steady-state condition did not differ from one another, t (39) = .08. A time of presentation by type of irrelevant speech interaction emerged, F (1, 38) = 14.78, MSe = 0.01. It is clear from Figure 1 the irrelevant speech effects were stronger in the rehearsal phase condition than in the input phase condition. Importantly, however, simple effects analyses indicated that irrelevant speech effects were evident in both conditions. Thus, in the input phase performance under the quiet condition was significantly greater than under the changing-state, t (19) = 3.59, and the steady-state condition, t (19) = 2.36. Correct recall for steady-state and changing-state irrelevant speech did not differ from one another under the input conditions t (19) = 1.00. For the rehearsal period performance was significantly superior under the quiet condition than steady-state, t (19) = 5.01, and changing-state conditions, t (19) = 7.02. Changing-state and steady-state conditions did not significantly differ from one another, t (19) = .69. The results are quite straightforward. Irrelevant speech has a detrimental impact upon cued recall just as it does on immediate serial recall and a number of other memory tasks. Secondly, the effects of irrelevant speech are more pronounced when the auditory input occurs during a rehearsal period rather than during presentation. This result is somewhat surprising given that more irrelevant speech is experienced when the auditory material is presented during study than when it is presented in a rehearsal period. It might be the case that even though subjects are requested to rehearse during the rehearsal period, some forgetting is taking place. For instance on the quiet condition, recall after a two second unfilled retention interval is not as good as on an immediate test. Furthermore, with irrelevant speech it is possible that the auditory material effectively makes this rehearsal period a filled retention Feature Interactions 16 interval. Tolan and Tehan (1999) have shown that very brief periods of verbal distractor activity produce a marked negative impact upon the cued recall task. Alternatively, from an O-OER perspective, irrelevant speech effects are expected to be produced where subjects are relying upon order information. In the delayed test students are instructed to rehearse the item whereas no such instructions were given for the immediate test group. To the extent that subjects are more reliant upon rehearsal in the delayed condition than in the immediate condition, irrelevant speech effects should be more prominent after a delay. Finally, and perhaps most importantly, both steady-state and changing-state irrelevant speech appear to produce equivalent levels of disruption and this is true irrespective of when the irrelevant speech is presented. This finding is atypical in that changing-state speech usually produces a robust decrement, but steady-state speech produces either a weak effect or no effect at all. Since the current results are somewhat unexpected we thought it best to replicate the findings before placing too much weight on the current results. Thus in the next experiment we again manipulate the type of irrelevant speech we employ. The effects of irrelevant speech on proactive interference are also examined. Experiment 2 Experiment 1 clearly demonstrated the effects of irrelevant speech under cued recall procedures. From a feature interaction perspective the irrelevant speech has its effect by becoming incorporated into memory. Here we assume that the primary effect of the irrelevant speech is on the phonemic representations in memory. Thus, the phonemic features of list items will be embedded within a background that contains the phonemic features of the irrelevant items. Tehan and Humphreys (1995, 1996; 1998) have argued that the phonemic representations of list items serve as a means of producing immunity to PI. Thus, if these Feature Interactions 17 representations are degraded by the irrelevant speech, their ability to assist in isolating the most recent memories from earlier memories may well be reduced. Thus, in the current experiment we revert to the PI version of the cued recall task. On some trials an interfering foil is presented in the first block. We expect that under quiet conditions, the presence of the foil will have little impact because the phonemic features of the target are retrieved against a background of minimal trace degradation and these features will reinforce the target. That is, we expect to observe immunity to PI. However, if irrelevant speech has an impact upon the phonemic features of the target, it is possible that target recall will become more difficult and that discriminating between the target and the foil will be more difficult. PI should be observed under irrelevant speech conditions. We again use both steady-state and changing-state irrelevant speech, although in this experiment it is manipulated between subjects. Furthermore, the irrelevant speech was only presented during a two-second rehearsal period. Experiment 1 indicated that placement of the irrelevant speech during the rehearsal phase was a more potent manipulation even though the patterns performance were identical for both methods of presenting the auditory material. Methods Participants A total of sixty participants took part in this experiment and were randomly assigned to one of three group according to order of arrival (n = 20 per group). The groups differed on the type of irrelevant speech they experienced. Methods The selection of materials and list construction was similar to that used in Experiment 1. This time, two instances from forty taxonomic categories from the South Florida norms were selected. Again following Tehan and Humphreys (1995) the target item tended to be a low dominant item from with in the category and the second which served as the foil in the Feature Interactions 18 interference trials, was a relatively high dominant instance from within the category, although never the most dominant. The filler items were selected from the same sources and care was taken so that there was no category overlap between the fillers and the target/foils. Fifty-two lists, of which forty were two-block trials and twelve were one-block trials were constructed in the same way as in Experiment 1, save that on half the two-block trials, one of the filler items in the first block was replaced by the foil. The foil and the target were always allocated to the same serial position in their respective blocks; half the time in the second position and half the time in position three. Again each subject was given a unique set of trials in which the materials had been randomly allocated to the different trials and the order of the trials had been randomised as well. The irrelevant speech items were the same as those used in Experiment 1. Procedure The procedure was identical to Experiment 1 except that irrelevant speech was manipulated between subjects. In the changing-state and steady-state conditions two groups of twenty participants were each exposed to irrelevant speech during a two-second rehearsal phase that appeared immediately after the fourth TBR item in the second bock of words and before presentation of the category cue. For the quiet condition the irrelevant speech was absent during the two-second rehearsal phase. It was, however, emphasised in all conditions that the two-second retention interval was to be used to rehearse the four TBR items and that the concurrent irrelevant speech should be ignored. Results Figure 2. presents the data for target recall and the errors made. PI effects can be determined either by differences in target recall or by the change in block-1 intrusions. Both recall measures were analysed. Feature Interactions 19 Correct recall The correct recall data was submitted to a 2 x 3 mixed design analysis of variance, with interference being a within-subject manipulation and irrelevant speech a between-subjects factor. A significant main effect for irrelevant speech was obtained, F (2, 57) = 5.90, MSe = 0.02. Planned comparisons revealed that recall of the target item was significantly better under quiet conditions than steady-state, t (38) = 2.33, and changing-state, t (38) = 3.57, conditions. Target recall for changing-state and steady-state conditions did not significantly differ from one another, t (38) = 1.03. The main effect for interference was not significant, F (1, 57) = .57, MSe = 0.01, confirming that there was no difference in correct recall between interference and non-interference conditions. In addition, there was no significant interaction between irrelevant speech and interference, F (2,57) = 1.51, MSe = 0.01. Block-one intrusions Planned comparisons indicated that in that significantly more foils were produced under the changing-state condition compared with the quiet condition, t (38) = 2.57. The steady-state condition also produced significantly more block-one intrusions than the quiet condition, t (38) = 2.19. The steady-state and changing-state conditions did not significantly differ from one another in the amount of block-one intrusions produced, t (38) = .14. Discussion Recall performance for the target item showed an irrelevant speech effect. Furthermore, steady-state and changing-state speech appeared to produce roughly equivalent levels of interference. There were some signs that changing-state speech resulted in more omission errors than steady-state speech, but in Experiment 1 this was not the case. Likewise, there were no reliable differences in block-1 intrusions between the two forms of irrelevant speech. Thus, the results of Experiment 2 replicate those of Experiment 1 by showing that Feature Interactions 20 there is an irrelevant speech effect in the cued recall task and that steady-state and changingstate speech produce equivalent levels of disruption. The above description of the differences between steady-state and changing-state effects is based on instance by instance comparisons. It is possible to do an experiment wide meta-analysis of the results as well. We first calculated the effect size for the difference between the quiet and the steady-state means for all conditions in both experiments. Effect sizes based on the mean differences between quite and changing-states were also calculated and all effects sizes are reported in Table 1. Inspection of Table 1 suggests that the effects of changing-state speech are stronger than the effects of steady-state speech in all conditions. This difference was statistically reliable, t (8) = 7.26, p < .000. Hence this further analysis of the combined data supports the notion that changing-state irrelevant material does have a greater disruptive effect on memory performance than steady-state irrelevant material, although obviously the differential effects are much weaker in the cued recall task than they are in immediate serial recall. We explore this issue further in the General Discussion. With regards to PI effects in the experiment, there was no significant difference between interference and non-interference conditions for the correct recall data. But this appears to be the result of different error patterns. That is, while the total number of errors are roughly equivalent in the three conditions, the frequency of the different types of errors change. Thus, more omission errors are made in the no-interference conditions than the interference conditions and this is offset by increased block-1 errors in the interference conditions. Given that block-one intrusions are the strongest measure of PI effects, the data indicate that irrelevant speech results in recall being vulnerable to PI. The speech interferes with the processes that produce immunity to PI in the cued recall task. There is substantial evidence that irrelevant speech eliminates the phonological similarity effect (Colle & Welsh, 1976; Jones & Macken, 1995; Salamé & Baddeley, 1982). This suggests that irrelevant speech Feature Interactions 21 interferes with phonological codes to the extent that the available phonological information is no longer of a sufficient quality that it unambiguously supports the target item.