There are many roads to Rome, none of which can be considered "right" or "best". I did a bit of testing on this, using this snippet by maestro Brian Eno from SampleSwap.
I actually tried to work this out with Audacity, but it crashed on me, so I reverted back to Ocenaudio, which has been my weapon of choice for this kind of work for almost 2 decades.
I opened the file in both Audacity and Ocenaudio, split it to mono files and saved one of the mono files to new wavs. I then exported the mono file from Ocenaudio with different settings:
- 320, 160 and 96 kbps
- Fixed Bitrate
- Variable Bitrate with qualities Highest, Medium and Lowest (7 in total available)
The resulting file sizes:
The original audio is a 44.1KHz 16bit stereo WAV file.
Both Audacity and Ocenaudio work internally with 32bit floating point data, and Audacity's default export is with that encoding; hence the large file size. It can also export with other encodings but I didn't test due crash. Ocenaudio's 32bit export is slightly smaller, and dropping the bitrate to 16bit halves the size.
All the MP3 files are exported from the 16bit mono file. All of these files have - to my ear - perfectly fine quality when the point is just to keep the speech clear and understandable.
For the two smallest files I first converted the original 44.1K mono file to 8K, then exported it as 8 kbps MP3, FBR and VBR. These have a significant drop in the audio quality, but only with the smallest file bad enough to really bother me.
These results give an indication of what can be expected, but these aren't set to stone. These values are valid for this specific piece of audio using this specific version of Ocenaudio. Different source material and / or application is likely to produce somewhat different results.
Assuming the target is a portable device with average in-ear headphones, my approach to the job would be something like this:
- Split the file to mono
- Manipulate the mono file for maximum clarity. This usually includes:
- Normalize the file to -3db - -6dB. I might go directly to 0dB as well, as the other steps somewhat reduce the level
- Apply a 12dB high pass filter at 120 - 200 Hz to remove low-end rumble and plosives (low-freq pops produced by p, k, t etc.)
- Apply EQ. I generally put a 3dB - 6dB boost somewhere between 1000K and 4000K range, depending on the source material (eg. male voice clarity generally sits a bit lower than a female voice). Sometimes it also involves making a 3dB or so dip somewhere around 500Hz - 800Hz. I often also introduce a low-pass filter at 8000K - 10000K.
- If the recording has noise, I'd introduce noise reduction at this point
- Apply compression, maybe 2:1 - 6:1 with threshold around -3dB
- Export the file as MP3 with the desired setting
In the end the only way to find the acceptable balance is simply testing.
If I had to work a bunch of files my workflow would be similar, but I wouldn't use either Audacity or Ocenaudio. I'd use a proper multitrack tool which allows me to load the same set of FX with the same settings to each track, bounce them in place and export all files at one go. With Ocenaudio and AFAIK Audacity the FX must be applied to each file separately, one by one.