Speech test recordings - where to obtain?

[A.S.]

I've been asked by a prospective user "what would make a suitable speech test recording .... old radio shows?". It's a very good question. I've written about the critical importance of using clean, well recorded speech as a vital speaker evaluation tool.

Is anyone interested in me making available on the Harbeth website a selection of what I consider to be useful speech recordings that I have made myself with details of their known limitations? For example, one I have in mind is a little bit 'bathroomy' because I recorded in in a less than ideal acoustic, but I am aware of this and ignore this recording characteristic when comparing speakers.

[Don Leman]

I would certainly appreciate you doing this for the group.

Don

[TNIC]

I agree this would be a useful download.

As for MP3 or WAV we (the downloader?s) could down load either format and use the available tools to convert them to another format ourselves if that is necessary.

To prevent endless support questions on reformatting, both formats can be posted and the downloader?s could choose.

Thank you for your thoughtfulness.
TNic

[A.S.]

Don,

Here is my first (snipped) audio test piece. I have used it extensively over the past years as I paid the announcer to buy-out the copyright. The mic is a U87.

Chesty: possibly but equally likely that it is reflections off the woodwork around the announcer. A similar problem to the floor bounce effect with speakers.

Regarding the BBC, I regret that today's broadcast studio environment seems to emphasis modern architectural trends by using glass rather than damped cloth surfaces. In addition, sonically hard, 'attention-grabbing' microphones are in vogue as are reverberant acoustic spaces which add a definite twang to the announcers voice between the words.

Here is how BBC announcer, Brian Empringham sounded in a well damped, now abandoned, studio at Bush House, London in 1989. Microphone was U87. Picture here of very similar studio to the one in which this was recorded.

Loading the player ...

By the way: this analogue 1/4 inch recording is not perfect. You can hear tape print-through (unavoidable with analogue tape) but the character of the voice is clean enough to allow side by side comparisons of loudspeakers for which it is good enough.

[A.S.]

Don Leman in Canada commented that he was aware of some chestiness in the broadcast voice of Canadian Broadcasting Corporation (CBC) announcers. I noted that the acoustic environment of the broadcast studio had changed markedly over the years both in the BBC and throughout the industry as server-playout systems have become the norm.

In today's all-digital studio environment more directional microphones are used, there is far less sound proofing on the walls (cost and floor-load weight reduction?) and there are many more hard reflective equipment surfaces. DJs/presenters need to have TFT (touch) screens within easy reach as the entire playout systems are now completely computerised: real CDs are rarely handled in today's broadcast studio. TFT screens have created a vertical wall behind and around the microphone which is now at the concave focal point of the array of vertical screens - not acoustically ideal. This contrasts with the lower operating desk height of previous generation of presenter's studios.

Here are some pictures including those of the late and much missed Peel, Moore and Dunn that illustrate the analogue-era studio (acoustic) environment up to today's all-digital situation. Note in the older pictures the proximity of record racks and cartridge stacks which by accident (?) acted as mini sound diffusers minimising the contribution of reflections behind and under the presenter's voice. The studio environment now has a very different reverberation characteristic.

What is of particular interest to me is that in a couple of these picture the Harwood-BBC designed 4038 ribbon microphone is quite clear. It has a figure-of-8 polar pickup pattern which made it ideal for a face-to-face interview, since one microphone would pick-up both parties equally loud. Note how far the presenter is from the 4038 - perhaps 500mm or so - and contrast that with today's presenter whose lips are almost touching the microphone's windshield. However, when the 4038 was used as a single presenter's mic it must have been prone to pick-up the sonic signature of the room. That we never heard or were aware of the room acoustic implies to me that the acoustics environment in the studios in the 1950s through 70s was much kinder - and doubtless very much more expensive to achieve. Overkill?

Also note that the BBC had an original policy of designing their in-house mixing desks such that individual channel faders were fully on when pulled towards the presenter, and fully off when pushed away. This was (and is) the reverse of the industry-wide norm but was quite deliberate: if a presenter accidentally knocked a fader he would most likely push it away and hence reduce the level. Eventually, the industry won, and the 'reverse logic' BBC channel faders gave way to convention of louder if pushed away. It always amazes me how sound engineers in the BBC can work in different studios some with older desks with reverse faders and still manage to mix programmes correctly! It looks so confusing.

If the programme is live you can see the current studio webcam here: http://www.bbc.co.uk/radio2/ and click the link 'webcam' on the top right of the page. If there is no link, then the programme may be recorded, and no webcam is available.

Conclusion:

You probably should not use current DJ/presenter/announcer's voices off-air as audio test material. All radio station managers create a 'station sound' by manipulating the audio (microphone, studio acoustics, compression, frequency response shaping) to hold an audience. It may or may not be wise to use radio dramas. It is always best to make your own trusty recordings outside on a windless day with the best microphone you can borrow.

[A.S.]

It may or may not be good enough to use as a 'reference' test source. It depends upon ...

1. Studio acoustics
2. Type of microphone and where it is placed relative to the performer
3. The performers vocal ability. It is surprising how 'reedy' some voices are when listened to close-up.
4. How much post-processing; equalisation, dynamic range compression etc. etc.
5. Limitations of the recording, editing, duplication process.

Assuming the studio is all-digital from the mic-input through to the CD then I'd say that the most worrying would be items 2 and 4.

I will try and make time to look out some recordings that I've made. For highest fidelity the microphone should really be an omni or figure-8 which rules out all voice-booth type recordings since they would, almost by definition in a small space, be cardiod with that particular sonic signature.

[A.S.]

I have booked the BBC's facility for measuring speakers. Derek, who is coming along to help, has had a great idea: after the measuring (and assuming that we have time) we will record our voices to DAT tape using the same reference microphone.

We will edit them down into a suitable format and make them available for download.

[A.S.]

When we were at the BBC anechoic chamber recently Derek Hughes and I made a speech recording onto DAT (16 bit, 48kHz) from the B&K 4191 reference microphone that we used to measure the speakers. Because we are in an anechoic chamber there is no reflection so our voices are as 'pure' as possible. I played this recording to Walter Swanbon our USA distributor when he visited. He found them to be 'eerily natural' on the Monitor 40.1 prototype.

If anyone is interested in this recording of us chatting about speakers then I will edit it and make it available. It would provide a useful reference for speech under near perfect conditions whether or not you know our voices well. To minimise the file size it would have to be an MP3 file. I would convert from WAV to MP3 at a high bit rate.

[Will]

I'd be very interested in getting a copy of this file/recording.

[Knut Knutsler]

Hi Alan

I think this would be an exceedingly interesting recording with which to experiment. Please do make it available.

[shseto]

oh, yes please. This is a great idea !

[A.S.]

This is a mono recording.

Here is our informal conversation about 'What makes a good test recording'. We were in the BBC anechoic chamber using the same B&K microphone that we used to measure our speakers. The recording was direct to DAT (=uncompressed, linear). The B&K preamplifier or head amplifier is a little hissy but that's because it is a measurement microphone not a studio microphone where low noise is traded for super-flatness. We were about 0.7m from the omnidirectional microphone capsule.

The B&K preamplifier was the BBC's: the B&K amplifier and B&K capsule were Harbeths.

This is a high quality MP3/Pro file. Ideally, you should play to one speaker alone. Due to the nature of listening in real rooms, you should expect some differences when playing back a mono recording on either the left or right speaker alone which even themselves out when listening to stereo recordings.

Picture of exchanging the BBC supplied but uncalibrated (1960s) BBC microphone capsule for Harbeth brand new B&K 4191 (made 2007) calibrated capsule before raising the motorised mic boom to be level with our mouths. Please note, the script that we have read-out is a specially designed combination of words that barely make sense but provides a special energy balance. The B&K data sheet is attached and you can see how extremely flat the microphone is (the capsule is supplied with an individual response curve).

Revised file link, please go to post #16

[joel]

I think you should both do more recordings. An extremely accomplished and informative double-act! I couldn't help thinking how much this sounded like the beeb used to sound - even through my little Stax headphones (the "portable" ones - pretty good for what they are).
One thing, though. It is stated with great confidence that speech came before music, however Steven Mithen begs to disagree... Now even if you disagree with Mithen, this is a thought-provoking read.

[ellison]

An old thread, I know, but I'm new here. If you ever do a speech test V3, how about a female voice, too? She might excite (or demonstrate absence of) a different band of speaker resonances! But thanks for the demo: this clean almost-unprocessed clip is very useful to me as a home-studio male narrator, since it's hard to evaluate anything on one's own voice.

[Pharos]

Alan, You may remember that I asked you for advice on obtaining a reference speech recoding for assessing loudspeaker performance some months ago, and my suggesting doing a stereo recording in room, my hoping to reproduce the acoustic of the venue room in the listening room. You advised against this and suggested that I record externally to try to achieve as near as possible an anechoic recording.

Racked by some of the excruciatingly poor TV sound this evening, I have been forced into an apprehensive doubt about my speaker performance, my getting the impression that there is often little energy above 2kHz in close conversations on many TV programmes. R4s 'Today' programme was also very poor over the Christmas period, but this may have been compounded by the recent return to Broadcasting House.

I at that time had no idea about the presence of this thread, and would now very much like to use your second edited recording for listening, but the link appears faulty, and on clicking 'here' the site returns me to the 'here' position. Is it possible please to restore this accessibility.

[A.S.]

Alan, You may remember that.... I at that time had no idea about the presence of this thread, and would now very much like to use your second edited recording for listening, but the link appears faulty, and on clicking 'here' the site returns me to the 'here' position. Is it possible please to restore this accessibility...

You're absolutely correct. I've fixed the links and tidied this thred.

Luckily, the V2 source file is still safely on the new server.

You can download it here. If you right click on it you can save it as a high quality mono MP3 file.

Now that I am fully committed to developing the subject of how we actually hear on HUG and trying to put it on a rational basis as far as judging audio equipment is concerned, I will find more examples of speech that could be used as analytical tools. My Studer 810 was recently repaired (our old adversary, leaking PSU capacitors to blame) and I can transcribe some more old BBC analogue tapes.

[A.S.]

Well known BBC Radio 3 Announcer, Cormac Rigby talks about some experiences during his time there when received pronunciation really mattered.

Loading the player ...

Again, this may not be a technically perfect recording, but it is good enough for the purposes of identifying differences between loudspeakers.

[A.S.]

Another Announcer's voice from the past at BBC Radio 3: Patricia Hughes Also here. And here "that dreadful woman...". I did not realise until posting this item with the audio clip I've had in my archive for some years that sadly, Patricia died yesterday

Loading the player ...

Attached image shows the sonic spectrum of the three BBC announcers v. the recording of Derek and myself. As you can see, all four recordings have a very similar gradient although they were recorded with different microphones, different equipment, different studios and many years apart. Brian Empringham was certainly recorded on 1/4 analogue tape and Derek and I was recorded direct to digital.

>

[Pharos]

Thank you Alan for restoring the availability of your Derek and Alan recording. I have not listened to it on my system yet having not yet burnt it to CD because I have been busy with arduous work on the house, I have briefly listened on my PC speakers.

Musing on the smooth and extended top end through the sibilants, and also what seemed to be rather a lot of bass given my PC speaker's probable response, I came to the conclusion that the sibilant spectrum between about 5 and 8K was probably that of a set of histogram spikes descending down smoothly in amplitude.

That is exactly what your spectrogram shows, a fairly gently downward slope towards the top end; but I am also surprised at the amount of bass shown in all the recordings, even below 100Hz on speech.

The clarity of the enunciation is very marked, and it represents the importance of RP at the time.

This aspect has concerned me in that current pronunciation, which is very much less controlled, may actually place a much greater demand on a loudspeaker. Often the mouth and voice are not now used as controlling modulators, but as imparters of sets of explosive and percussive sounds, which tend to grab attention, I think because that is how aggression is expressed.

This style also often, because of lack of disciplined control, results in the vocal chords being allowed to 'flap', producing a croak or squark after the percussive sound, rather, as mentioned before, like the springs on a snare drum, which I think is particularly so with the female voice. Also, subjectively perceived, American males tend to produce percussive bass with very little elsewhere in the spectrum, this rendering it devoid of some phonemes and relatively unintelligible.

Perhaps the 'new' uncontrolled voice is a greater test of the loudspeaker.

[A.S.]

It occurred to me overnight that as we have noted that five voices (include Derek and myself) have a very similar energy gradient with frequency, we could define our own random noise profile which averages the spectral characteristics of the voices and applies that average to random noise. Let's call this new noise profile 'H noise'*, human noise.

If you are curious about how a loudspeaker reproduces human voice but you do not have a reliable speech test source you could arrange an A-B comparison whereby listening only to H noise would give you a direct comparison of the energy distribution across the loudspeaker's audio band, and excessive top should be readily apparent on listening and being a characteristic of the speaker not the human voice. We could say that H noise is a minimum standard for speaker energy balance because we know that id human speech is not reproduced naturally, music cannot sound natural.

So, starting with white noise, the random noise you would hear between channels on an FM radio, surface noise of an LP record, tape hiss or similar broadband noise it sounds like this ....

Loading the player ...

Unprocessed white noise

If we take that random noise and we apply a gentle filter to progressively reduce loudness as frequency increases we have generated pink noise. Pink noise has far less HF content and is generally used in room acoustics because it is much less likely to overheat and damage tweeters (and much less fatiguing to listen to as you will discover).

Loading the player ...

Pink noise filter applied to white noise

I examined the downward sloping spectral trend-line of the very high quality digital recording (made to DAT) of Derek and my voice in the anechoic chamber (previous post) captured with a truly flat microphone, and devised an audio filter which bent the white (or further filtered the pink noise) spectrum until it was very close to the voice spectrum. I allowed a little more output above 10kHz. See attached graph. So, H noise sounds like this:

Loading the player ...

Processed pink noise = H noise = averaged speech

As you can see from the graph, at 100Hz there is the same loudness from our voices and H noise. H noise sounds quieter than pink noise, and much quieter than white noise, because there is so much less 'top' in voice than in white or pink noise, which H noise mimics. The essential point is that in the real world, the amount of energy in speech and music rapidly diminishes as frequency increases, so a super-extended frequency response is most unlikely to be relevant or audible. Were it not for the fact that there is proportionately so little HF in speech and music relative to the mid and low frequencies, analogue recording and production would be impossible: the stylus simply wouldn't be able to accelerate fast enough to trace high level, high frequency signals: the G force would be destructive.

For background understanding of how to interpret frequency response charts, please refer here.

It also follows that designing audio equipment for a ruler-flat 20Hz-20kHz audio bandwidth where it is assumed in the lab that the signal can be full power at either ends of the audio spectrum (as white noise is) is not a worthy design goal. As we know from the BBC's long experience, a brickwall filter at 15kHz in the studio deprives the serious classical listener of nothing useful.

* 'H noise' © Harbeth Audio Ltd.

>