Studio versus field speech: design an acoustic evaluation Compare capture conditions and processing history for speech models without universal SNR thresholds or fixed clean/noisy training ratios. Includes a reproducible inspection example. Studio and field recordings describe capture contexts, not automatic quality grades. A controlled microphone source may suit a clean-voice task while failing to represent a phone or distant-room deployment. Field recordings can preserve relevant conditions but also contain defects or poorly documented labels. Choose sources through matched evaluation and a documented preparation path, rather than impose a universal signal-to-noise cutoff or a fixed percentage of noisy audio. Choose by Task and deployment conditions Keep distinct Capture condition, content difficulty and file defects Preserve Originals and processing history Experiment Controlled source-condition comparison Define the conditions in terms an engineer can inspect “Studio” may mean a treated recording room, a home microphone setup or simply a polished release. “In the wild” may mean a phone, a wearable, a distant microphone or a remote-call export. Those labels are useful starting points, but they omit device, channel, environment, editing and processing details. Write the relevant conditions explicitly. A car-cabin dictation product, a conference-room assistant and a synthetic narrator do not need the same source mix. Inspect which aspects of the target interaction must remain authentic and which controlled variations can be introduced during training. Separate defects from difficult but valid conditions Background conversation, reverberation or a narrow telephone channel may be part of the deployment. Missing samples, an unreadable file or a clipped target word may be defects that prevent a useful reference. The acceptance policy should distinguish these instead of calling everything outside a quiet microphone recording “low quality.” Also distinguish perceptual preference from model suitability. Denoising can sound pleasant while changing quiet speech or timing. A normalized export can be easier to ingest while concealing how the source was recorded. Keep the original and describe the transformation so the effect can be tested. Inspect file metadata before making acoustic claims [FFmpeg’s ffprobe documentation](https://ffmpeg.org/ffprobe.html) describes machine-readable container and stream inspection. The command below is an example for a local file. It reads metadata; it does not modify the source, measure recognition quality or establish that a recording is truly studio-captured. ffprobe -v error -select_streams a:0 \ -show_entries stream=codec_name,sample_rate,channels,channel_layout:format=duration \ -of json source.wav Interpret metadata without overstating what it proves A 48kHz file can be a higher-rate export of an originally narrow-band recording. Two channels can be a duplicate stereo mix rather than independent participants. A lossless container can hold audio previously compressed by a lossy codec. Those possibilities are why source and processing history belong beside measured file properties. The [Transformers ASR guide](https://huggingface.co/docs/transformers/en/tasks/asr) illustrates resampling to the selected model’s expected input. Keep that preparation step separate from the capture record. A derived 16kHz input is not evidence that every original was captured at 16kHz or has the same acoustic bandwidth. An illustrative acoustic comparison This authored experiment is a decision aid, not a claimed result. Suppose a product must recognize delivery instructions spoken into phones inside vehicles. Build a fixed, appropriately held-out evaluation set representing that intended condition, then compare carefully scoped training changes. Training condition Control to keep stable Decision it can inform Existing close-microphone source Baseline checkpoint, training budget and text policy How far the current source transfers Close-microphone source with documented augmentation Same source identities and evaluation set Whether the chosen perturbation addresses the observed gap Additional authentic target-condition recordings Comparable annotation policy and tracked source counts Whether new capture adds useful evidence beyond augmentation Enhanced derivative of target recordings Original/derived linkage and matching evaluation policy Whether the processing helps the task or removes useful information Read research results at their actual scope [SpecAugment](https://research.google/pubs/specaugment-a-simple-augmentation-method-for-automatic-speech-recognition/) evaluates feature-domain augmentation on named ASR benchmarks. It supports the idea of controlled perturbation as an experimental option, not a promise that synthetic noise reproduces every phone, cabin or interaction condition. Report the implementation, source relationships and conditions used in your own test. If the field recordings also contain different vocabulary or speakers, acknowledge that the comparison changes more than acoustics. A measured gain may still be useful, but it should not be described as a clean causal estimate of “studio versus field.” A review sheet for recording and processing history Capture a small number of useful fields consistently. Unknowns are preferable to inferred values presented as measurements. The aim is to make source differences interpretable when a model succeeds or fails. Record source device/channel where supported and note unknown capture settings. Keep sample rate, channels, codec and duration as measured file properties. Distinguish raw tracks, mixes, edited exports and enhanced derivatives. Record gain changes, resampling, filtering and other known transformations. Review clipping, missing material and transcript/audio mismatch separately from legitimate noise. Keep source families and related transformations in the same evaluation group. Choose the next source from the error review A fixed rule such as “use 30% noisy audio” cannot identify what your model lacks. Review whether errors arise from distance, background speech, channel, vocabulary, annotation or another condition. Select the smallest next experiment that can distinguish the plausible causes. Fiund can discuss a source and preparation scope against that requirement. A sample should demonstrate the relevant recording context and expose limitations; a category label or polished waveform should not substitute for that evidence. Availability, technical acceptance and permitted uses need to be confirmed for the proposed material. What SNR makes audio good enough for training? There is no universal cutoff established here. Define the task and measurement method, inspect actual defects, and evaluate the conditions the product must handle. A source can be difficult yet appropriate for robustness work. Can added noise replace target-condition recordings? It can test a particular perturbation, but it may not reproduce real channels, behavior or interactions. Compare the intervention on suitable held-out target speech rather than assume equivalence. Should enhanced audio replace the original? Keep the original and link the enhanced version as a derivative. Test the processing against the task; listening pleasantness alone is not proof that speech content or timing was preserved. Research method: Fiund research desk; public sources, explicit worked-example labels and project-specific availability review. Editorial contact: hello@fiund.com. Sources ffprobe documentation: https://ffmpeg.org/ffprobe.html Transformers automatic speech recognition task guide: https://huggingface.co/docs/transformers/en/tasks/asr SpecAugment: A Simple Augmentation Method for Automatic Speech Recognition: https://research.google/pubs/specaugment-a-simple-augmentation-method-for-automatic-speech-recognition/