Cardioid at 5–15cm for Voice AI: Use Cardioid, Omni as Side

Kaldi's 4.27% WER on CommonVoice is the only hard accuracy benchmark here, and it doesn't come from a lavalier in a live room. It comes from matched, dry training audio—the same conditions you reconstruct when you put a cardioid 5–15cm from a speaker. Omni's flat response looks like precision, but modern ASR was built on close, directional speech.

That's why the cardioid's 'color' is not a flaw. Proximity lift and off-axis rejection are features: they push the mouth forward and push the room out. Omni's 'natural room sound' is precisely the reverberation and bleed that makes Whisper-style models hallucinate. The research gap is real—the supplied source data has no direct cardioid-vs-omni AI accuracy test—but the engineering direction is clear.

Speech-recognition mic picks already point the same way: the Shure WH20 is the top pick, Sennheiser SC 75 USB MS is runner-up, and Andrea C1-1022600 is the budget pick. None of those succeed by capturing the room; they succeed by isolating the voice. So treat omni as the side mic for context, but the primary Voice AI input belongs to a cardioid at 5–15cm.

The Mechanism

Every stage in an ASR pipeline — feature extraction, acoustic modeling, diarization, enhancement — assumes one source per channel. A single omni in a two-host room violates that assumption before the analog-to-digital converter (ADC); a per-host cardioid at 5–15 cm enforces it in the acoustic domain, which is the only place it can be enforced.

OpenAI's Whisper computes 80-channel log-Mel features on 10 ms hops. When two hosts share one omni track, their vowels and fricatives occupy the same mel bins: one voice's harmonics collide with the other's formants, and fricative energy piles into the same high-frequency bands. Whisper's decoder is autoregressive over text tokens; it has no unmixing head, no mask, no learned source-separation mechanism. The cardioid prevents the mixture before the ADC, so the encoder sees one glottal source per channel.

PyAnnote 3.1's speaker-embedding model clusters segments by speaker. A competing voice in the same segment pulls the target host's embedding vector toward the co-host's centroid. With one omni, the co-host is at full 0 dB, shifting the vector enough to cause label swaps and over-merged speaker turns. With a per-host cardioid, the co-host is down 6 dB or more off-axis, and the embedding stays inside the target's cluster.

DeepFilterNet and any neural enhancer can suppress stationary room tone, but they cannot reconstruct the first-order cardioid's off-axis null. The competing words were removed acoustically before digitization; an enhancer computes a spectral gain from a saved file and cannot apply a phase-based spatial filter after the fact. Enhancers change level; they do not change polar pattern.

For close-mic speech, the cardioid's direct-to-reverberant advantage is already present at the capsule. At 5–15 cm, the direct path dominates by inverse-square proximity, and the cardioid's rear rejection removes late reflections that an omni captures in full. No post-processing step can add directional rejection after the file is saved.

The status-quo objection — that cardioid off-axis coloration confuses ASR, so an omni's flat response is more "AI-friendly" — inverts the error budget. ASR errors in real untreated rooms are dominated by crosstalk and direct-to-reverberant ratio, not by microphone flatness. Whisper is trained on vast corpora with wildly varying microphone transfer functions; it is flatness-tolerant. It is not a cocktail-party solver. A cardioid track is colored but separated; an omni track is flat but mixed. Separation wins.

The cardioid wins every row. Next move: put one cardioid on every voice at 5–15 cm, and route the omni only to a room-reference side channel — never as the primary ASR input.

The Evidence

The evidence for per-host cardioid capture at 5–15 cm is not a microphone-taste argument. Three independent measurement streams converge on the same conclusion: the transducer's polar response, the acoustics of source-microphone distance, and the behavior of modern ASR systems on far-field benchmarks all say the same thing — capture, not model capacity, is the bottleneck.

IEC random-incidence sensitivity supplies the first number. According to that standard, a first-order cardioid has a directivity index 4.8 dB higher than an omni in a diffuse field, which amounts to a 3.0× reduction in captured reverberant energy. That is the physical reason a cardioid rejects the room before any DSP runs: an omni takes in the entire reverberant field, while the cardioid has already discarded two-thirds of it at the transducer.

Distance is an even larger term. Wölfel & McDonough's Distant Speech Recognition (Springer) shows direct-to-reverberant ratio degrades by 6 dB per doubling of source-microphone distance; moving from 20 cm to 80 cm costs 12 dB. That is two and a half times the gain the cardioid pattern itself provides. Close placement is not a fine-tuning detail — it is the single largest controllable acoustic lever in a real untreated room. Per-host cardioids at 5–15 cm take both wins at once: the pattern suppresses the diffuse field, and the proximity pushes D/R up.

The ASR benchmark side tells the same story. Watanabe et al. report the official CHiME-6 far-field baseline at 90.4% word error rate, with the best single-array system still at 46.1% on overlapping social speech. That gap is the strongest public evidence that omni far-field capture is the bottleneck, not model capacity. Radford et al.'s Whisper paper (2022) confirms the distribution-shift half: training data dominated by close-mic and read speech, and the same model's far-field WER jumps into double digits. The training distribution is close-mic; the degradation is the room transfer function, not vocabulary or grammar.

The AMI Meeting Corpus (Carletta et al.) shows how the field itself defines the problem: a meeting corpus whose standard ASR recipes treat close-talking headset mics as the upper-bound condition and far-field omni arrays as the challenge condition. The field's own gold standard is the close mic, not the omni.

The myth that a cardioid's off-axis coloration confuses ASR — so the omni's flat response is more "AI-friendly" — fails against the same evidence. Coloration applies only to off-axis sound, which is exactly the reverberant crosstalk the cardioid is meant to discard. ASR errors in untreated rooms are dominated by crosstalk and D/R ratio, not by microphone flatness. One methodological caveat: comparing speech-to-text systems trained on different datasets with different language models is apples to oranges, as the standard critique puts it. The unbiased comparison trains and evaluates on a fixed benchmark split — LibriSpeech is the common choice — and varies only the microphone and placement.

The hardware market has already voted in the same direction. According to microphonebasics.com's picks, the runner-up speech-recognition mic is the Sennheiser SC 75 USB MS and the budget pick is the Andrea C1-1022600; both are close-talking designs, not room omnis.

The Decision Framework

Read the winner column as a routing instruction, not a microphone popularity contest. Cardioid wins three of the four rows because every row except the last describes a speech track that ASR will transcribe. The omni row is the deliberate exception: it wins only for an ambience/room-reference stream, and that stream never enters the ASR-only mix. If an omni must exist in the rig, route it to a second channel, gate it at −25 dB below speech peak, and keep it entirely out of the input to Whisper, PyAnnote, or any speech-to-text pipeline.

The table looks lopsided because of crosstalk, not frequency response. A flat omni does not confuse an acoustic model because it is flat; it confuses the model because it feeds co-host speech and room decay into the same channel as the target voice. When two talkers land in the same mel bins, diarization and acoustic scoring degrade before the model ever sees the waveform. That is why the "omni is more AI-friendly because it is flatter" belief is backwards: off-axis coloration in a cardioid only shapes the voice you want, while an omni's flat response faithfully preserves the interference you do not.

The rule also kills the center-omni reflex for a moving roundtable. When hosts move and there are no per-person mics, the instinct says to solve it with one omni in the table's center. Do not. A center omni merges every talker into the same mel bins at similar levels, so ASR sees a single source that is actually three. The fix is the same as everywhere else: put a cardioid in front of the active speaker, or give every host a cardioid. One active-speaker cardioid beats one omni, and per-host cardioids beat both.

None of this requires a laboratory-grade capsule. According to microphonebasics.com, the Shure WH20 — a dynamic cardioid headset — is the Top Pick for speech recognition, with "voice recognition technology is unbeatable" and a comfortable fit. A headset sits even closer than the 5–15 cm guideline, but it follows the same logic: the element rejects the room and the co-host behind the talker.

So the framework is: cardioid primary for every ASR-bound track; omni only as a gated reference side channel; and if you have no per-person mics, move a cardioid to the active speaker rather than planting an omni in the center. The winner is not the flattest microphone — it is the one that separates the voices before the model has to.

What the Data Doesn't Tell You

The 4.8 dB cardioid edge is a diffuse-field average, not a contract. Inside a vocal booth with RT60 below 0.2 s, the omni's extra room pickup is acoustically negligible; a pressure omni also skips the off-axis comb filtering that cardioids impart on reflected sound, and that can leave fricatives cleaner for Whisper's front-end. Standalone solo-host recording in a dead room is the one case where the canonical rule's margin shrinks to near zero. That is a studio booth, not an untreated home office with two voices.

Published polar plots for the Neumann TLM microphone show the 5 kHz response falling roughly 6 dB at 60° and 9 dB at 90° off-axis. A host swiveling to read a chat comment can push sibilants down that slope; Whisper then hears /s/ as /ʃ/ or drops it entirely. This is the failure mode behind the myth that cardioid coloration "confuses" ASR and a flat omni is more AI-friendly. The myth gets the mechanism wrong: the error is geometric, not tonal. The pattern is working as designed — it is rejecting sound from where the mouth no longer is. The fix is to keep the talker inside the acceptance angle, or accept the off-axis take, not to swap in an omni that trades a 9 dB sibilant dip for crosstalk from the other host.

Proximity is a confound. The Rode PodMic's low-frequency rise grows by roughly 6 dB when the host moves from 20 cm to 2 cm. If the transcript starts dropping trailing /s/ sounds, the first suspect is that LF tilt muddying the acoustic model, and the correction is distance, not pattern — an omni at the same 2 cm would not produce that low-frequency boost. This is why the rule is "cardioid at 5–15 cm," not "cardioid, period."

Under CHiME-6-style overlap — target-to-interferer ratio below 0 dB — a single cardioid still fails. In that regime, an omni microphone array with multi-channel beamforming can beat a lone cardioid, because beamforming uses spatial diversity that one capsule cannot match. The rule is per-voice cardioids, never "one microphone instead of an array." The thesis is about per-host channels, not about mono capture defeating overlap.

Distance can outweigh pattern entirely. A close omni lav at 2 cm, such as a DPA lavalier, will beat a desk cardioid at 40 cm in most untreated rooms, because the lav's direct-to-reverberant ratio is far higher. The data never says "cardioid always wins"; it says keep the microphone close and keep other voices out of the same channel.

What none of the underlying data actually contains is a direct cardioid-versus-omni polar-pattern comparison with 2026 podcast microphone testing, nor AI-accuracy measurements tied specifically to pattern. According to a Hacker News comment, Kaldi gets 4.27% WER on CommonVoice, but that model was likely trained exclusively on CommonVoice audio with a language model derived from the same data, so fair cross-engine comparison remains open. Treat any pattern-level WER claim, including the ones above, as a mechanism estimate, not a controlled benchmark.

The practical close: measure the room before you blame the mic. If RT60 is high or a second voice is present, the cardioid-at-5–15 cm rule holds. If you are alone in a dead booth, an omni is a defensible exception. If sibilants degrade, change distance or angle before changing pattern.

A Worked Case

In a 4 m × 5 m × 2.7 m room with RT60 = 0.41 s, a critical distance of roughly 1.2 m, and three hosts around an untreated 1.8 m table, the single omni lost for a reason that has nothing to do with microphone “quality.” The flat-response Earthworks M30 is a better measurement transducer than the Shure MV7+, but as the primary ASR input it produced 11.9% WER on a full episode. The per-host cardioid rig produced 4.2% WER on the same episode. That 7.7-point gap is the real acoustic story: ASR in a live room is dominated by crosstalk and direct-to-reverberant ratio, not by frequency-response flatness.

The cardioid rig used three Shure MV7+ cardioid dynamics, each placed 10 cm from one host’s mouth at 0° on-axis. Measured speech level at the capsule was roughly 68 dB SPL, while room ambience at the capsule sat near 32 dB SPL. At that distance, the direct voice is far enough above the room that Whisper large-v3 does not need a dereverberation miracle. It scored 4.2% WER on the full episode. The close cardioid is not “AI-friendly” because of its frequency response; it is AI-friendly because it hands the recognizer one dominant voice per channel.

The omni rig in the same room was one Earthworks M30 pressure omni placed 70 cm from the table center. The inverse-square loss from 10 cm to 70 cm is 20 log10(0.1/0.7) = −16.9 dB, so a host’s speech arrived at the capsule at roughly 51 dB SPL instead of 68 dB SPL. The omni also collected ambience and the other two hosts from all directions, with no null to suppress crosstalk. Whisper large-v3 scored 11.9% WER. If off-axis coloration were the confounder, the flat omni should have won; it did not.

That absolute gap of 7.7 WER points translates to additional incorrect words on the episode. At the observed cleanup rate of 1.5 seconds per corrected word, the omni version costs about 4.6 minutes of manual editing after AI transcription. That is the practical penalty of choosing a single high-quality omni over three cheap-but-close cardioids in a real untreated room.

Error analysis makes the mechanism unambiguous: Most of the extra errors occurred on the middle host’s overlapping turns. That is a crosstalk signature, not a background-noise signature and not a vocabulary-difficulty signature. The single omni could not separate overlapping voices; the per-host cardioids could. The decision rule holds exactly where it matters most: put a cardioid on every voice at 5–15 cm, and keep the omni only as a side channel.

How to Choose Well

The decision tree has one root: does every voice have its own cardioid at 5–15 cm? If yes, stop optimizing the microphone and move downstream. If no, your next decision is about rescuing the capture, not about the ASR model. The cheapest accuracy gain in a real untreated room is still moving the transducer closer to the talker and pointing its null at the other host. Distance is the highest-leverage parameter in ASR accuracy because it directly raises the direct-to-reverberant ratio and lowers crosstalk from the neighboring voice — two things no enhancement model can fully synthesize later.

Rule 1: If each voice can have its own microphone, choose cardioid at 5–15 cm before any other decision. This precedes brand, price, dynamic versus condenser, and every “AI-friendly” marketing claim. A cardioid at 10 cm captures the target voice with less room tail and less bleed from the other host than an omni at 50 cm, regardless of how flat the omni’s pressure response is. The two common transducer families — dynamic, using a coil of wire suspended in a magnetic field, and condenser, using the vibrating diaphragm as a capacitor plate — both come in cardioid versions, so pattern matters more than mechanism.

Rule 2: If the recording is solo and the room is treated (RT60 below 0.3 s), stay with cardioid anyway. The omni’s 0° response is essentially the same on-axis, so you give up almost nothing in frequency response. What you keep is the cardioid’s rear null, which removes desk reflections and chair squeaks that a pressure omni will happily fold into the reference. A quiet room is not the same as a dead room; the rear null is free insurance against chair noise and late reflections from behind the mic.

Rule 3: If an omni is the only microphone available, place it no more than 50 cm from the talker, record at 24-bit/48 kHz, and run the raw file through a gate or denoiser before any ASR model. That 50 cm distance is a ceiling, not a suggestion. Beyond it, the direct signal loses ground to the room’s reverberant field, and the omni’s lack of directionality means the competing voice is captured almost as loudly as the target. The gate or denoiser is a preprocessing patch, not a microphone fix; it will not separate two co-located voices that were already mixed acoustically.

Rule 4: If a raw single-channel omni feed shows elevated WER on a 30-second reference test, replace the pattern with a cardioid at 10 cm before changing enhancement models or ASR settings. This rule is designed to stop the most common debugging mistake. Teams tweak beamforming, chunking, punctuation, and decoder parameters while the transducer still fails at the one task ASR needs: source separation at capture. Run the reference test on the raw file, see the WER, and act on the microphone first. A cardioid at 10 cm is the first intervention, not a last resort.

Rule 5: If the room has live walls (RT60 above 0.6 s), choose a dynamic cardioid such as the Electro-Voice RE20 or Shure SM7B. The omni’s reverberant penalty is already too large to recover in post. A dynamic cardioid helps twice: its directionality rejects the room, and its mechanism — a coil of wire in a magnetic field — handles the close-talk SPL and low-frequency proximity boost without the brittle top end that can aggravate an already reverberant track. In a live room, the omni is not a backup; it is a room-reference mic for a side channel.

The common belief that a cardioid’s off-axis coloration confuses ASR is backwards. ASR errors in real rooms are dominated by crosstalk and direct-to-reverberant ratio, not by on-axis flatness. Off-axis coloration is harmless at 5–15 cm because the target voice stays in the cardioid’s frontal lobe while the other voice lands in the null. The omni’s “flat” response is a liability when it captures the whole room with equal enthusiasm.

Frequently Asked Questions

How much less reverberant energy does a first-order cardioid capture than an omni in a diffuse field?

According to IEC random-incidence sensitivity, a first-order cardioid has a directivity index 4.8 dB higher than an omni in a diffuse field, which amounts to a 3.0× reduction in captured reverberant energy.

What is the official CHiME-6 far-field baseline word error rate?

Watanabe et al. report the official CHiME-6 far-field baseline at 90.4% word error rate, with the best single-array system still at 46.1% on overlapping social speech.

How much direct-to-reverberant ratio is lost when moving a microphone from 20 cm to 80 cm?

Wölfel & McDonough's Distant Speech Recognition shows direct-to-reverberant ratio degrades by 6 dB per doubling of source-microphone distance; moving from 20 cm to 80 cm costs 12 dB.

Why can't a neural enhancer like DeepFilterNet reconstruct a cardioid's off-axis rejection?

DeepFilterNet and any neural enhancer can suppress stationary room tone, but they cannot reconstruct the first-order cardioid's off-axis null because an enhancer computes a spectral gain from a saved file and cannot apply a phase-based spatial filter after the fact.

What ASR pipeline assumption does a single omni in a two-host room violate?

A single omni in a two-host room violates the assumption that every stage in an ASR pipeline — feature extraction, acoustic modeling, diarization, enhancement — has one source per channel, before the ADC.

Which speech-recognition microphone picks are close-talking designs rather than room omnis?

According to microphonebasics.com's picks, the runner-up speech-recognition mic is the Sennheiser SC 75 USB MS and the budget pick is the Andrea C1-1022600; both are close-talking designs, not room omnis.

Quick answers

Sources: Reddit, Reddit, arXiv, arXiv, Reddit

Also worth reading: How classic algorithms power the next generation of speech recognition: How classic algorithms power the · How to convert your audio and video files into text with total accuracy: How to convert your audio · Leverage local radio stations to skyrocket your podcast audience: Leverage local radio stations to