Voice Across Contexts¶

Conversational speech, recorded speech, telephone calls, public presentations, and singing all make different demands on the same underlying voice. A voice that holds together in casual conversation can collapse on a phone call where high-frequency formant cues are filtered out. A voice that records well at room volume can be unrecognizable in a noisy restaurant. Training in one context doesn't automatically transfer to another.

This page covers the major context-specific demands and the adjustments each requires.

The unifying principle: voice training produces a target acoustic output (fundamental frequency, formant frequencies, resonance, prosody, breath). Each listening context applies a different transfer function to that output — filtering some cues, amplifying others, adding latency, raising the noise floor, demanding louder production. A voice tuned for one context is not the same voice as the one a listener actually hears in another.

Conversational speech (the baseline)¶

Conversational speech is the context most learners train in: a quiet room, roughly one meter of listener distance, full audio bandwidth (about 80 Hz to 8+ kHz of perceptually relevant content), no transmission filtering, minimal latency, and visual cues available. Almost every published acoustic target — F0 around 165–185 Hz, brighter resonance, lifted /i/ and /e/ formants — assumes this context.

All the perceptual gender cues are intact in conversation:

F0 (fundamental frequency) — the pitch of the voice.
F1, F2, F3 (formants) — resonance and vowel quality.
Prosody — sentence melody, stress, phrase contour.
Voice quality — breathiness, glottal source quality, register.
Articulation detail — sibilants, fricatives, vowel precision.

Listeners also have non-acoustic information: face, body, clothing, context. The voice does not have to do all the work alone. This is part of why in-person voice tends to read better than the same voice on a phone call: the visual channel is doing some of the gender-perception work.

The implication for training: metrics derived from conversational practice are calibrated for the conversational context. They are not automatically valid in other contexts. A voice that hits target F0 and resonance brightness in front of a desk microphone in a quiet room may need different settings on a phone, in a loud room, or while projecting to the back of a hall. The rest of this page is about which adjustments transfer and which do not.

Recorded voice¶

Why your recorded voice sounds different¶

When you speak, you hear yourself through two paths simultaneously:

Air conduction — sound traveling out of your mouth, around your head, and into your ear canal.
Bone conduction — vibration through the skull and jaw directly to the inner ear. Bone conduction emphasizes low frequencies, because skull bone acts as a low-pass filter relative to the air path.

Recorded playback contains only the air-conduction path. The low-frequency emphasis you are used to hearing is gone. The recorded voice therefore sounds thinner, higher, and brighter than the internally heard voice.

This is not the recording lying. The recording is closer to what other people actually hear. Internal hearing is the version that is biased.

For trans women specifically, this bias usually runs against the learner: internal hearing makes the voice sound deeper and less feminine than it reads externally. Recordings often sound "more feminine than I felt," which is uncomfortable in a different way — some learners report dissociation or cringe — but the underlying signal is favorable. The recording is the external-listener's-eye view, and external listeners are who matters.

The mic-feedback discomfort¶

The gap between internal perception and external recording is one of the highest sources of training-avoidance. Many learners describe physical discomfort hearing themselves played back, especially in the first weeks.

The discomfort is real but the recording is the only objective measurement available without paid coaching. The internal monitor cannot be trusted — it is filtering the very signal you are trying to evaluate.

Practical framing:

Recordings are non-negotiable for progress. You cannot train what you cannot hear.
The discomfort fades with exposure. Most learners report it diminishing substantially over 2–6 weeks of regular listening.
Early sessions can be short. Sixty seconds of speech with a single playback is enough to start, and short sessions reduce overload.
A voice coach effectively serves the same function — an external ear reporting back what the internal monitor cannot. Self-recording is the budget version of the same feedback loop.

For app inventory, recording technique, and longitudinal logging, see Recording & Self-Monitoring.

Phone microphones, video calls, voicemail¶

Each common recording chain has a distinct acoustic profile and degrades the signal differently. A voice that sounds good on built-in laptop playback may not sound good through the same phone the doctor's office hears.

Chain	Bandwidth	Typical issues
Phone mic, handheld call	Narrowband (PSTN)	F3 stripped, codec compression, AGC, noise suppression
Smartphone speakerphone	Wider than PSTN	Far-field mic; reverberation and background noise
Laptop / built-in mic	Wideband-capable	Distance from mouth, room reverb, fan noise
Video call (Opus / SILK)	16 kHz+ when ideal	Aggressive denoising; breathy voices misread as noise
External USB / lavalier mic	Full bandwidth	Closer to studio reference; least common in daily life
Voicemail	Narrowband	Stored through telephony path regardless of capture

Two practical implications:

Voicemail is a phone-context task even if you are recording into a smartphone with a wideband mic. The storage path goes through telephony infrastructure and clamps the audio to PSTN bandwidth on the way through.
Video-call denoising is tuned to typical cisgender voices. Aggressive noise suppression and "voice isolation" features can mangle breathy or quiet feminine production, classifying it as background noise. If your voice on Zoom sounds choppy, gated, or unnaturally clipped between syllables, try turning voice isolation off and letting the platform send the signal through more transparently.

The minimum useful self-recording protocol is to capture yourself across all the chains you actually use — phone call to a friend who can record their side, a Zoom call recorded locally, a voicemail you leave for yourself. Compare the chains; the gaps between them are the gaps your training has to close.

Telephone voice — the highest-anxiety context¶

The phone is the context most learners report as the largest gap between in-person and remote read. Misgendering on the phone after months of successful in-person presentation is a common, demoralizing experience. The mechanism is not that your training is failing. It is that the phone literally removes information your training depends on.

Why the phone is harder¶

Narrowband filtering (300–3400 Hz)¶

Traditional public-switched-telephone-network (PSTN) telephony specifies a 300–3400 Hz audio band. This is codified in ITU-T G.711, the µ-law / A-law codec that has been the backbone of telephony since the early 1970s. Anything below 300 Hz or above 3400 Hz is filtered out before the signal is transmitted. Even fully digital paths emulate this band for backward compatibility.

Wideband VoIP and HD voice extend this band substantially:

Codec / standard	Approximate band	Where it appears
ITU-T G.711 (PSTN)	300–3400 Hz	Landlines, voicemail, most carrier bridging
ITU-T G.722	50–7000 Hz	HD voice on capable carriers
3GPP AMR-WB	50–7000 Hz	Mobile-to-mobile HD voice (VoLTE, VoNR)
IETF Opus / SILK	50–8000 Hz+	WebRTC, Discord, Meet, Zoom, Teams

Most carrier-to-carrier calls still drop to narrowband at some point in the chain because at least one leg is PSTN or legacy mobile. End-to-end HD voice requires both endpoints and the entire path to support it. In practice:

Calls to a business landline, a doctor's office, a government agency, an older PBX, or any toll-free routing — assume narrowband.
Mobile-to-mobile on the same carrier with both endpoints on a recent network — possibly wideband.
App-to-app over the internet (Signal, WhatsApp, FaceTime audio) — usually wideband.

The high-stakes calls (medical, government, customer service) are the ones most likely to be narrowband. The casual calls with friends are the ones most likely to have HD voice. Training instinct should default to narrowband for anything that matters.

Formant cues above 3.4 kHz are stripped¶

Resonance cues to perceived gender live across the full vocal spectrum, but they are not evenly distributed:

F1 (typically 300–900 Hz) — survives the narrowband filter intact.
F2 (typically 800–2500 Hz) — survives intact.
F3 (typically 2300–3500 Hz for adults, often higher in trained feminine voices) — sits right at the upper edge of the narrowband filter and is partially or fully stripped.
F4 / F5 and higher spectral energy — gone.

The higher spectral energy that contributes to perceived brightness, "head resonance," and the s/f/sh sibilant clarity associated with feminine speech is largely above 3.4 kHz. Strip those cues and a voice with feminine F0 can still read as masculine, because the remaining cues (F1, F2, F0, prosody) are not by themselves sufficient for the listener.

This is the mechanism behind "I sound great in person but get sirred on the phone." It is not a training failure. It is the channel doing what the channel does. Earlier work by Spencer (1988), Gelfer & Schofield (2000), and later Wolfe and colleagues established that F3 and the higher formants carry a substantial portion of the listener gender-perception signal independent of F0 — and the phone is, by design, a device that throws those cues away.

For the underlying acoustics of formants and what they encode, see Resonance Fundamentals.

Latency disrupts feedback¶

Voice production relies on rapid auditory feedback to the speaker — your own voice, in your own ear, closing the audio-motor control loop in tens of milliseconds. The phone disrupts this loop in two ways:

Round-trip latency. Phone calls have round-trip latency typically 150–400 ms (mobile-to-mobile is often higher; international or VoIP-bridged calls higher still). This is long enough to disrupt real-time calibration of resonance and pitch.
Sidetone variance. Sidetone is the small amount of your own voice fed back into the earpiece on many phones. It helps close the loop. It is not always present, especially on speakerphone, earbud calls, or calls routed through certain Bluetooth profiles.

Compensating: rely more on internal proprioceptive cues — jaw position, tongue position, breath, the felt forward placement of resonance — rather than on the auditory monitor. Your "feel" of the voice has to lead, because the auditory loop is impaired. Learners who have done dedicated forward- placement work (see Resonance Fundamentals) have an easier time on the phone for this reason.

Compensating on the phone¶

The strategy is not to push harder on every dimension. It is to lean into the cues that survive the filter and lean off the cues that don't.

Lean harder on:

F1 and F2 — vowel quality. More deliberate vowel shaping, especially on /i/, /e/, and /æ/, which carry strong gendered formant differences.
Articulation precision — crisper consonants, longer voiceless fricatives (/s/, /f/), avoid mumbling. Sibilant placement matters even more on the phone than in person, because the sibilant brightness is partly preserved while higher resonance is not. See Articulation.
Intonation and prosody — wider pitch range, more melodic contour, upward inflection on phrase ends where it is natural. Prosody is fully preserved by the phone channel and carries real gender-perception weight. See Intonation & Prosody.

Lean less on:

Resonance brightness above ~3 kHz. Trying harder here produces strain without improving the listener's signal. The energy you generate above 3.4 kHz is being thrown away before it reaches the listener.
Breathy quality. Much of the high-frequency aspiration noise that carries the breathy character is filtered or mangled by noise suppression. Excessive breathiness on the phone often just sounds quiet.

F0 still matters. Keep it in target range. Do not try to compensate for missing F3 by raising F0 further. Pushing F0 above ~210 Hz to overcome a phone-clipped voice tends to sound effortful and reads as performed rather than feminine.

Speak slightly slower than in person. A slower rate gives the codec more time per phoneme and gives the listener more time to integrate the reduced-information signal.

The phone-rehearsal trick

Record yourself reading a script into your phone, save it as a voice memo, then call your own voicemail and leave the same script. Listen to both. The voicemail playback is what the doctor's office actually hears. The gap between the two is the gap your phone-context training has to close.

Scripts for common high-stakes calls¶

The cognitive-load tradeoff: a fully scripted call has lower cognitive load during speaking, which leaves more attention for voice production. A natural unscripted call has higher cognitive load, which tends to collapse trained voice features back toward the pre-training default. Most learners benefit from scripted openers and key phrases for high-stakes calls, then natural speech for the back-and-forth once the call is established.

The first 10–15 seconds set the gender read for the whole call. Spend the preparation budget there.

Medical office¶

Information you will be asked for: name, date of birth, reason for visit, insurance carrier and member ID, sometimes a callback number. Have all of it written down before you dial.

Pre-script the opening:

"Hi, this is [Full Name], calling to confirm my appointment on [Date] / reschedule the appointment I have on [Date] / ask about [reason]."

Anticipate the verification block:

"Date of birth is [DOB]. The insurance is [Carrier], member ID [Number]."

If gender markers on the chart do not match your presentation, decide before the call whether to address it explicitly ("My chart still has the old name listed; I go by [Name] now") or let the verification land and move on. Both are reasonable; the former front-loads the awkwardness, the latter shortens the call.

Government calls combine the longest holds with the most adversarial back-and-forth. Pre-script the request and have all reference numbers written down before you dial.

Expect to be asked to verify identity multiple times, sometimes by reading back digits or spelling your name. Practice saying your full legal name in target voice ahead of time. Names are short, high-attention, and often the exact moment a call goes wrong.

Pre-script the request:

"Hi, I'm calling about [case number / claim number / application]. My name is [Full Name], [DOB]. I'm trying to [specific request]."

If the agent misgenders you, the question is whether to correct it. There is no universal answer. In most cases the call will go faster if you let it pass and focus on completing the transaction; the federal employee on the other end is rarely the right venue for the correction. If a name or marker on the record is actively blocking the request, address that record, not the agent's pronouns.

For long government holds, voice fatigue is real. Take a sip of water at any hold. If voice is slipping past 30 minutes on hold, end and call back later; a tired voice on the second hour does not survive the verification block.

Customer service¶

Customer service is easier than government because the agent is incentivized to be cooperative and the script options are usually broader. The risk is the call going long — voice fatigue after 20+ minutes is real, and a tired voice slips. Take water at any hold.

Pre-script the opening only:

"Hi, I'm calling about [issue]. The account is under [Name], [account > number

/ phone number on file]."

Do not try to push through fatigue on a long support call. If voice is slipping, ask for a callback (most modern call centers offer this) or hang up and call back fresh. The transaction will go faster with a rested voice than with a tired one fighting through verification.

When the voice slips mid-call¶

If you feel the voice dropping during the call:

Pause. A two-second silence is fine; agents pause too.
Take a small breath, drop the jaw, re-find forward placement.
Resume on a short, vowel-forward phrase ("Yes, that's right" / "Okay, so —").
Do not announce the slip. Do not try to "fix" the past five seconds. The listener has already moved on.

Public speaking and presenting¶

Voice with a microphone¶

A microphone carries the volume. The speaker's job is no longer to project — it is to articulate clearly, vary prosody, and pace. Trying to project into a microphone is the most common amateur error, and it produces a voice that reads strained without being any louder.

Breath support shifts from powering loudness to powering steady airflow for clear voicing. Underused breath produces a weak, breathy mic voice that picks up every plosive and click. See Breath Support.

Mic technique:

Distance: typically 4–8 inches for a handheld, closer for a podcast-style condenser, further for a stage shotgun.
Off-axis: turning the head off-axis from the mic drops level and changes tone. If you have to turn, move the mic with you.
Plosives: /p/ and /b/ blast the diaphragm of the capsule. A pop filter helps; angling the mic 15–30° off-axis to the mouth helps more.
Lapel and headset mics pick up close-in detail including breathing and swallowing. Practice with the actual mic type if possible.

Bad mic technique can undo good voice training. A learner with a feminine voice on a poorly placed lapel mic can sound boomy, popped, and harsh through the PA — none of which is the voice's fault.

Voice without a microphone¶

Projection has to come from breath support and resonance, not from squeezing the throat. Throat-driven loudness pushes F0 up and back, often producing a strained voice that reads more masculine, not less. The mechanism: increased subglottic pressure with a closed-down resonator elevates pitch but also recruits the laryngeal musculature in a way that darkens the timbre.

Demands on F0 stability are higher than in conversational or phone speech. A long unamplified speech is closer to athletic performance than to everyday voice use. Many learners find projection is the last skill to come online — well after conversational voice is solid.

Plan accordingly:

Avoid unamplified public speaking early in training if you can.
For unavoidable unamplified work, prepare with breath-support drills, not pitch drills. The bottleneck is air, not pitch.
Voice rest before and water during are not optional for projection-heavy speaking. Caffeine and dairy beforehand are usually counterproductive (drying / mucus, respectively).
If a venue has any amplification at all, use it. There is no prize for doing it raw.

Singing voice¶

Speaking range vs singing range¶

Speaking voice and singing voice are distinct skill sets that draw on overlapping but not identical mechanics. A trained speaking voice does not automatically transfer to singing.

Dimension	Speaking	Singing
Pitch range used	1/3 to 1/2 octave around habitual F0	1.5 to 2+ octaves untrained, more with training
Register use	Mostly modal	Modal, mixed, head/falsetto, transitions between
Sustained tones	Rare (mostly transient pitches)	Constant; vibrato and steadiness matter
Breath demand	Conversational phrases	Long phrases at controlled pressure
Resonance strategy	Forward placement for gender perception	Resonance varies by register, repertoire, style

Trans women who want a singing voice should train it as a separate practice with a singing teacher, ideally one with trans-voice experience. Speaking-voice work alone will not deliver it. Conversely, singing training does not by itself produce a feminine speaking voice; the demands overlap but the targets differ.

For overall pitch theory, including the difference between habitual, optimal, and target speaking F0, see Pitch Fundamentals.

Pre- and post-HRT singing changes¶

Estrogen does not significantly remodel adult vocal folds. Vocal-fold length, mass, and the resulting fundamental-frequency range are largely set at the end of testosterone-driven puberty and are not reversed by HRT.

Singers on estrogen typically retain their pre-HRT singing range. Some report subjective changes in tone quality (often described as "softer" or "lighter") and changes in stamina; these are not well characterized in research and individual experience varies. Cycle-related water retention and mucosal effects can affect voice from day to day on cyclic regimens; these effects are minor in most singers.

The practical takeaway: HRT is not a singing-range intervention. Singers should not expect their range to migrate upward on estrogen the way puberty migrates a singing range downward on testosterone. The window where HRT might affect singing range — pre-puberty or peri-puberty suppression with later estrogen — is a different clinical situation than adult-onset feminization, and the long-term singing data there is extremely limited.

Pre- and post-VFS singing impact¶

Voice feminization surgery aims to raise speaking F0. The mechanisms — shortening or thinning the vibrating portion of the vocal folds, or approximating the cricoid and thyroid cartilages to increase tension — all affect singing as well as speaking. Singers considering VFS should understand that the surgery operates on the same instrument they sing with.

The two most-discussed procedures in this context:

Wendler's glottoplasty (and related anterior glottal-web procedures). The anterior portion of the vocal folds is sutured together, shortening the vibrating length. Wendler's is known to narrow singing range, sometimes substantially. Many surgeons counsel singers carefully and some decline to operate on professional singers without explicit informed consent. Loss of lower range is typical; loss of upper range, vibrato quality, or smooth register transitions is reported variably. See Wendler's Glottoplasty.
Cricothyroid approximation (CTA). The cricoid and thyroid cartilages are brought together to mimic CT muscle action, increasing fold tension. CTA is often described as having less impact on singing than glottoplasty, but it is also less reliable for sustained F0 elevation — many CTA results regress over months to years. See Cricothyroid Approximation.

Laser-assisted procedures (LAVA) and combined approaches have less long-term singing data published; outcomes appear closer to glottoplasty than to CTA, but cohorts are smaller.

Pre-op questions for a surgeon if singing matters¶

If singing is part of your life — professional, semi-professional, or just deeply personal — these are the questions to bring to the consult:

What is your experience operating on singers?
What singing-range outcomes have your singing patients reported?
How do you weigh F0 elevation against singing-range preservation?
Can you show me pre/post audio from a singer patient?
What is the realistic worst case for singing if I have this surgery?
For glottoplasty specifically: how do you decide on suture extent, and does that decision change for a singer?

A surgeon who has not thought carefully about the singer-specific tradeoff is the wrong surgeon for a singer. This is not about endorsing any particular practice — it is about whether the conversation happens at all.

When VFS narrows or shifts singing range¶

Some loss is expected with glottoplasty and may occur with other procedures. Plan for it.

Healing window. Singing teachers can help re-map range post-op once healing is complete — typically 6–12 months out, longer for full vibrato return. Trying to push range during the early healing window risks scarring and worse final outcomes. See VFS Recovery and Voice Training.
Repertoire shift. Keys may need to move. A song that sat in a comfortable middle range pre-op may need to be transposed up two or three semitones to clear the new lower limit, and the new upper limit may close off some material entirely.
Decline-or-defer. Some learners choose to delay or decline VFS specifically to preserve singing voice and rely on behavioral training plus, in some cases, resonance-only adjuncts. This is a defensible choice. The speaking gains from VFS are real, but so is the singing cost, and the tradeoff is personal.

For the broader VFS landscape, including which procedures fit which goals and how the candidacy decision is made, see VFS Overview.

Noisy environments¶

Restaurants, bars, public transit¶

The Lombard effect is the involuntary increase in vocal effort speakers produce in noise: louder volume, raised F0, slower rate, more vowel energy. It is reflexive and largely unconscious — first described by Étienne Lombard in 1911 and replicated extensively since. The Lombard reflex is not optional; the brain raises vocal effort whenever the ambient noise floor rises, regardless of what the speaker is consciously trying to do.

For trans women, the Lombard effect frequently blows out feminine voice training. The trained target sits at a particular volume and resonance; the brain in a loud room wants to produce something louder, which pulls F0 and resonance back toward the untrained baseline. Specifically:

Raised volume tends to drop the larynx (more chest resonance, deeper read).
Raised volume engages the vocal folds more heavily (rougher source, more pressed phonation).
Raised effort flattens prosody as more attention goes to projection and less to melodic contour.

All three run opposite to feminization targets. A learner who can hold target voice in a quiet café may find that voice collapsing in a busy restaurant without ever consciously deciding to push harder — the brain is doing it in the background.

Limited trans-specific Lombard research

Peer-reviewed work on the Lombard effect specifically in trans-feminine voice training is sparse. The general Lombard literature (Junqua and others) applies, but the specific magnitudes of larynx drop, F0 shift, and resonance loss in noisy environments for trained trans-feminine speakers are not well quantified. Treat the directionality as established and the magnitudes as personal — record yourself in the actual noisy contexts you care about.

When to compensate vs when to abandon¶

The decision is binary at the extremes and graded in the middle. Three rough zones:

Noise level	Strategy
Quiet café, quiet office, home	Train target voice; full feminization toolkit
Moderate restaurant, light street noise	Compensate: lean in, articulate, accept F0 lift
Bar at peak hours, subway platform, club	Abandon: step outside, lean close, or use text

When to compensate¶

In moderately noisy environments where the conversation matters:

Lean closer to the listener. Reducing distance is the single highest- leverage move. Half the distance, four times the apparent loudness — and no Lombard cost.
Use more articulation and prosody. Articulation and prosody survive volume increases better than resonance does.
Allow F0 to rise somewhat but consciously keep the larynx from dropping. The forward-placement work pays off here. If you have a felt sense of resonance position, you can hold it through a moderate volume increase.
Use closed-mouth posture between sentences. It conserves moisture, conserves breath, and conserves voice over a long meal.

When to abandon¶

When the environment is loud enough that even cisgender women's voices get strained — bars at peak hours, subway platforms, concert venues — trying to hold trained resonance at shouting volume produces strain, fatigue, and often a voice that reads worse than the untrained baseline.

Better strategies in these environments:

Step outside, or move to a quieter corner.
Lean in close enough to drop volume back into a sustainable range.
Use text for high-stakes information ("here's the address — texting it now").
Pick the quieter side of the table; sit with back to the room; ask to move tables; arrive early before the room fills.

Voice fatigue and recovery¶

Voice fatigue from environmental compensation is cumulative and recovers slowly. A two-hour loud dinner can leave the voice slipping for a day or two. Plan voice rest after high-Lombard situations:

Hydrate aggressively the next morning.
Avoid stacking loud-environment days back to back.
Do not schedule a high-stakes phone call for the morning after a loud night out.
Persistent hoarseness, pain, or loss of upper range past 48 hours is reason to stop voice training and rest, not to push harder. See Troubleshooting & Plateaus.

The non-voice-production compensations (table choice, seating, texting, arriving early) often matter more than any vocal technique. The best loud-restaurant strategy is the one that lets you not have to fight Lombard at all.

Sources¶

Voice training and trans-feminine voice¶

Davies S, Papp VG, Antoni C. Transfeminine voice and communication: vocal health and considerations for the speech-language pathologist. Perspectives on Voice and Voice Disorders 2015.
World Professional Association for Transgender Health. Standards of Care for the Health of Transgender and Gender Diverse People, Version 8 — voice and communication chapter.
Gelfer MP, Schofield KJ. Comparison of acoustic and perceptual measures of voice in male-to-female transsexuals perceived as female versus those perceived as male. Journal of Voice 2000.
Spencer LE. Speech characteristics of male-to-female transsexuals: a perceptual and acoustic study. Folia Phoniatrica 1988.
Wolfe VI, Ratusnik DL, Smith FH, Northrop G. Intonation and fundamental frequency in male-to-female transsexuals. Journal of Speech and Hearing Disorders 1990.

Telephony and codecs¶

ITU-T Recommendation G.711. Pulse code modulation (PCM) of voice frequencies — 300–3400 Hz narrowband PSTN specification.
ITU-T Recommendation G.722. 7 kHz audio-coding within 64 kbit/s — wideband HD voice.
3GPP TS 26.171. Adaptive Multi-Rate Wideband (AMR-WB) speech codec.
IETF RFC 6716. Definition of the Opus Audio Codec.

Lombard effect and noisy-environment speech¶

Lombard É. Le signe de l'élévation de la voix. Annales des Maladies de l'Oreille et du Larynx 1911. (Original description of the Lombard effect; foundational reference, no trans-specific application.)
Junqua J-C. The Lombard reflex and its role in human listeners and automatic speech recognizers. Journal of the Acoustical Society of America 1993. (General Lombard reflex literature.)

Peer-reviewed work on the Lombard effect specifically in trans-feminine voice training is sparse as of writing; treat magnitude claims as suggestive and verify against your own recordings in the contexts you care about.