MEMS Microphones and In-Ear Micro-Speakers for a G.729 Wireless Voice Link

Executive Summary

For the system you described, the easiest engineering path is not a literal implant, but an ear-canal, receiver-in-canal, or hearing-aid-style ear-worn assembly. On the input side, there are many strong CMOS MEMS microphone choices from Infineon, ST, and TDK. On the output side, the market is much narrower: the realistic published options are either miniature balanced-armature receivers from Knowles or solid-state/piezo MEMS speakers from xMEMS and USound. The former are the most mature for speech-first ear devices; the latter are attractive when you specifically want a magnet-free, solid-state transducer. None of the publicly documented parts I reviewed should be treated as implantable off the shelf for subdermal or middle-ear use, because implantable audio devices require long-term hermetic packaging, tissue-contact biocompatibility, and medical-device-specific design controls well beyond standard commercial MEMS packaging.

If your priority is speech intelligibility from a remote talker over G.729, the most important design variables are low microphone self-noise, low latency, strong output in the speech band, stable fit/sealing in the ear canal, and proper DSP around packet buffering, gain control, EQ, and limiting. In that speech-first use case, Knowles full-range balanced-armature receivers remain the most practical output path, while Infineon IM73A135V01, Infineon IM69D130, ST MP23DB02MM, and TDK ICS-43434 stand out on the microphone side. If you want the most “all-solid-state” receiver path, xMEMS Cowell is the strongest public MEMS-speaker candidate for hearing-aid/in-ear form factors, with Montara Plus as the higher-bandwidth but physically larger alternative.

G.729 itself is not a microphone-or-speaker hardware standard; it is a speech codec and packetization problem. In its basic mode, G.729 is an 8 kbit/s CS-ACELP speech codec using 10 ms frames with 5 ms look-ahead, and RTP commonly uses a 20 ms packetization interval, though 10 ms packets may also be used. That means your hardware chain should be thought of as G.729 decode → PCM/DSP → amplifier/ driver → ear transducer. Because the codec is narrowband speech-oriented, very wide playback bandwidth is less important than clean low-distortion output, low delay, and stable gain control.

System Context and G.729 Constraints

I am assuming your “USID on a G.729 wireless connection to a CPU” means there is a CPU or gateway endpoint that either encodes/decodes G.729 directly or passes G.729 frames over a wireless link to a receiver endpoint. In practical terms, any of the microphone chips below can feed a local ADC/PDM/I²S/TDM front end if you also want talkback, sidetone, voice activation, or ambient sensing. If your system is receive-only, then the local microphone is optional; the essential path is the decoder/DSP and ear transducer. This is an engineering inference from the codec and component interfaces described in the cited sources.

The most important G.729 engineering facts for your project are straightforward. At 8 kbit/s, one 10 ms G. 729 frame carries 80 bits, which is 10 bytes. With a 20 ms packetization time, payload becomes 20 bytes before link framing overhead. The codec’s minimum algorithmic delay is about 15 ms from frame plus look-ahead, and then transport, jitter buffering, decoding, resampling, and amplifier/rendering add more delay. ITU guidance on conversational delay notes that voice is affected by much lower delays than the absolute 400 ms upper planning bound, and earlier G.114 guidance identifies about 150 ms one-way as the region where interactivity is still generally acceptable. In practice, you want to keep packetization, jitter buffer, and render delay as small and deterministic as possible.

Because the transport is speech-oriented and bandwidth-efficient, your transducer selection should favor speech-band clarity, low distortion, and good output around the midband, not just extreme ultrasonic or hi-res extension. A MEMS speaker that reaches 40 kHz is technically impressive, but for basic G.729 voice it is not automatically more intelligible than a well-chosen balanced-armature receiver with strong spoken-word output. That conclusion follows from the G.729 narrowband design and the published receiver/speaker response classes in the component datasheets.

Remote talker microphone

CPU or gateway with G.729 encoder or decoder

Wireless link

Receiver MCU or DSP and jitter buffer

DAC or piezo/receiver driver amplifier

MEMS speaker or BA receiver

Ear canal or ear

Candidate MEMS Microphone Chips

On the microphone side, the design choice is mainly between analog and digital outputs. Analog microphones can minimize clock-domain complexity and sometimes offer very low power, but they need a clean analog front end and short routing. Digital microphones simplify routing into DSP/MCU silicon and are especially attractive when you want beamforming, ANC, or tight synchronization across multiple microphones. Infineon’s IM69D130 and TDK’s ICS-52000 are particularly strong when microphone matching or arrays matter; Infineon’s IM73A135V01 and ST’s MP23ABS1 are attractive where simple analog capture and low front-end current are more important.

Representative package classes are shown below. What matters physically is that modern MEMS microphones really are small enough for canal, concha, or receiver-in-canal shells. The receiver side, however, is more heterogeneous: true MEMS speakers are much fewer, and many practical ear-canal solutions still use balanced-armature receivers rather than a “CMOS MEMS speaker chip.”

Manufacturer

Part number

Interface

Package size

Sensitivity

SNR

Published

frequency response

Supply a current

Infineon

IM73A135V01

Differential analog

4.0 ×

3.0 ×

1.2 mm

-38 dBV typ

73 dB(A)

normal, 71 dB(A) low- power

20 Hz low- frequency cutoff; acoustic table published to 15 kHz; test bandwidth 20 Hz–20 kHz

170 µA t

normal,

µA typ lo power

Infineon

IM69D130

PDM

digital

4.0 ×

3.0 ×

1.2 mm

-36 dBFS typ

69 dB(A)

at 3.072 MHz;

lower- power modes available

28 Hz low- frequency cutoff; response table published to 15 kHz

Approx. 980 µA t dependi on clock/ mode

TDK InvenSense

ICS-43434

I²S digital

3.5 ×

2.65 × 0.98 mm

-26 dBFS ±1

dB

64 dBA

Wideband response with integrated high-pass corner at 24 Hz

Product literature publishe low-pow and high perform modes; about 23

490 µA

dependi on mode

Top Microphone Candidates

Manufacturer

Part number

Interface

Package size

Sensitivity

SNR

Published

frequency response

Supply a current

TDK InvenSense

ICS-52000

TDM

digital

4.0 ×

3.0 ×

1.0 mm

-26 dBFS

65 dBA

50 Hz to 20 kHz

1.0 mA t

TDK InvenSense

ICS-41350

PDM

digital

3.5 ×

2.65 × 0.98 mm

-26 dBFS in low-power/ standard,

-32 dBFS in high- performance mode

63–64 dBA

depending on mode

50 Hz to >20 kHz; ultrasonic support to 40 kHz in high- performance mode

185 µA l

power, 4

µA stand 650 µA h perform sleep bel 20 µA

STMicroelectronics

MP23DB02MM

PDM

digital

3.5 ×

2.65 × 0.98 mm

-26 dBFS ±1

64 dB(A)

low-power,

65 dB(A)

normal

35 Hz roll-off

at -3 dB; normal- mode audio bandwidth characterized to 20 kHz

2 µA slee

285 µA l

power, 8

µA norm

STMicroelectronics

MP34DT06J

PDM

digital

3.0 ×

4.0 ×

1.0 mm

-26 dBFS ±1

64 dB(A)

Datasheet publishes normalized frequency- response curve for audio-band use

650 µA t

µA powe down

STMicroelectronics

MP23ABS1

Single- ended analog

3.5 ×

2.65 × 0.98 mm

-38 dBV ±1

64 dB(A)

ST publishes frequency- response mask and describes it as ultra-flat

150 µA m

The top microphone conclusion is simple. If you want the best analog ear-worn mic in this survey, IM73A135V01 is the strongest published option because it combines 73 dB(A) SNR, very low current, and IP57 ingress robustness. If you want the best digital mic for arrays, beamforming, or tightly controlled DSP, IM69D130 is the most compelling because it explicitly publishes phase matching and array-oriented behavior. If your receiver MCU naturally wants I²S, the ICS-43434 is uniquely convenient; if you want ultra- low wake-current, MP23DB02MM is especially attractive.

Candidate Micro-Speakers and Tiny Transducers

The receiver/output side is where your question becomes much harder. There are many CMOS MEMS microphone chips, but there are far fewer true MEMS speaker chips suitable for deep ear-worn use. The public options I found split into two families. First are piezo/solid-state MEMS speakers from xMEMS and USound, which are physically small and magnet-free but require dedicated piezo drive electronics.

Second are balanced-armature receivers from Knowles, which are not MEMS chips but are still tiny transducers that dominate hearing-health and in-ear monitor designs because they are efficient, compact, and speech-friendly. xMEMS explicitly lists hearing aids among Cowell and Montara Plus applications, and Knowles’ hearing-health selection guide is explicitly aimed at traditional and OTC hearing devices.

Manufacturer

Part number

Type

Size

Published SPL

Frequency

response or role

Electrical load

Drive and power

Su

fo ca

xMEMS

Cowell

Full-range solid-state MEMS

microspeaker

3.2

× 6.0

× 1.15

mm

Current public page does not publish a numeric SPL;

launch coverage reported 110 dB

SPL at 1 kHz

20 Hz to

20 kHz in fully occluded systems; can also serve as a tweeter in 2-way modules

Piezo/solid- state load, not a simple coil impedance

Requires

xMEMS

Aptos2 amplifier path

B

ch M

sp he IE

cl

Top Speaker and Transducer Candidates

Manufacturer

Part number

Type

Size

Published SPL

Frequency

response or role

Electrical load

Drive and power

fo ca

xMEMS

Montara Plus

Full-range solid-state MEMS

microspeaker

5.15

× 10.8

× 1.15

Public page presents it as xMEMS’ highest- sensitivity MEMS

speaker; no public numeric SPL on current page

20 Hz to

>40 kHz

Piezo/solid- state load

Requires Aptos2 amplifier path

yo si dr M

an to la pa

xMEMS

Cypress

Full-range MEMS

speaker using “sound from ultrasound”

6.5

× 6.3

× 1.65

xMEMS

publicly claims

>140 dB

low- frequency SPL

Full-range ANC-

earbud- oriented speaker; optimized for low- frequency authority and ANC bandwidth

Ultrasound- modulated solid-state path

Requires dedicated xMEMS

ecosystem; current public press material pairs Cypress with Alta2

St ve fr ou m it na m sp ea th or re

USound

Achelous UT-P 2018

Piezo MEMS speaker

6.7

× 4.7

× 1.58

117 dB @

1 kHz / 15

Vp; 94 dB

@ 1 Vrms

Lower bandwidth

<20 Hz;

coupler SPL

published at 250 Hz,

1 kHz, 5 kHz

39 nF

capacitance

Up to 15 Vp, 200 mAp;

power around 13.6–

19.8 mW at published 94 dB test conditions

be do M

sp op in

Manufacturer

Part number

Type

Size

Published SPL

Frequency

response or role

Electrical load

Drive and power

fo ca

Knowles

RAQ-35091-000

Full-range balanced- armature receiver

5.3

× 2.7

× 1.6

109 dB

SPL / 100

mV, 117 dB SPL / 1 mW, 122

dB max

Full- range single BA; vented

74 Ω

Conventional receiver amplifier

Ex co sp op ea el in th M

Knowles

RLQ-34240-000

Dual- diaphragm full-range balanced- armature receiver

5.3

× 3.1

× 2.7

111 dB

SPL / 100

mV, 120 dB SPL / 1 mW, 126

dB max

Full- range, high- output

57 Ω

Conventional receiver amplifier

Ar th st al sp lo ch st ea pa

Knowles

BK-26824-000

Single full- range balanced- armature driver

7.9

× 5.6

× 4.0

114 dB

SPL / 100

mV, 118 dB SPL / 1 mW, 127

dB max

Single

full-range driver; can also be used as a woofer

22 Ω

Conventional receiver amplifier

Ve ou p la R

Knowles

ED-26805-000

Single full- range balanced- armature driver

6.3

× 4.3

× 3.0

109 dB

SPL / 100

mV, 113 dB SPL / 1 mW, 123

dB max

Single

full-range driver; Knowles also notes mid- range/ mid- tweeter use

28 Ω

Conventional receiver amplifier

pr m re co sy

For a pure “MEMS speaker only” interpretation of your request, the realistic shortlist collapses to Cowell, Montara Plus, Cypress, and Achelous. Among those, Cowell is the most naturally aligned to a hearing-aid or deep in-ear speech device; Montara Plus is better when you want wider bandwidth and can accept greater package size; Achelous is one of the most transparent public options because the datasheet actually publishes coupler SPL, capacitance, voltage, and power; and Cypress is technically impressive but more specialized and more integration-heavy.

For a practical speech-first design, however, the best answer is often a balanced-armature receiver, not a MEMS speaker. Knowles’ hearing-health and premium-audio parts were built specifically for small in-ear acoustic assemblies and can reach strong output from low-voltage electronics with simpler drive requirements than piezo MEMS. That is why they still remain the safer choice when the actual user requirement is “deliver intelligible voice into the ear canal.”

Power, Battery, Form Factor, and Implantation Feasibility

The microphone power budget is usually small compared with the overall ear device. Published microphone currents in the surveyed set range from the tens of microamps up to roughly 1 mA, depending on mode and interface. Output power is a bigger differentiator. A BA receiver with 109–120 dB class sensitivity can often be driven from a conventional low-voltage amplifier, while MEMS speaker approaches tend to require higher-voltage bias/drive stages. USound’s Achelous datasheet is unusually transparent here: it publishes 39 nF capacitance, 15 Vp maximum AC drive, 200 mAp maximum AC current, and roughly 13.6–19.8 mW power in its stated 94 dB test cases. xMEMS likewise makes public that Cowell and Montara Plus require the Aptos2 amplifier ecosystem, while USound publishes both Amalthea and Tarvos as MEMS-speaker drivers.

For battery selection, the most relevant published small-cell options are still hearing-aid and small wearable cells. Energizer size 312 zinc-air lists a typical capacity of 181 mAh, while size 675 lists 575 mAh. On the rechargeable side, the VARTA CoinPower family is a realistic custom-device option, with public material listing approximately 60 mAh for CP1254 A3, 85 mAh for CP1454 A3, and 120 mAh for CP1654 A3. As a rough design calculation, a 60 mAh rechargeable cell supports about 12 hours at 5 mA average or 6 hours at 10 mA average, while a 181 mAh zinc-air 312 cell supports about 36 hours at 5 mA average. Those runtime figures are simple current-capacity arithmetic, not manufacturer endurance claims.

For form factor, a custom ear-canal or receiver-in-canal shell is the realistic mechanical target. That is exactly the zone where Infineon, xMEMS, and Knowles all position relevant parts: Infineon’s IM73A135V01 explicitly targets ANC headphones and wireless earbuds, xMEMS lists hearing aids among Cowell and Montara Plus applications, and Knowles’ hearing-health receiver guide is explicitly for traditional and OTC hearing devices. In this class of device, ingress robustness, wax management, acoustic venting, and limiter/ EQ tuning become as important as the silicon part number itself.

A literal implant, by contrast, is a fundamentally different engineering problem. FDA biocompatibility guidance frames implantables around ISO 10993-style biological evaluation and risk management; the implantable-inner-ear MEMS packaging literature explicitly calls out biocompatibility and long-term hermeticity as core requirements; and FDA guidance for implantable middle-ear hearing devices treats them as premarket approval products. In other words, ordinary commercial MEMS microphone or speaker packages are not implant-ready simply because the die is small. If by “implant” you mean subdermal, middle-ear, or cochlear-area placement, none of the tabulated commercial parts should be considered appropriate without a medical-device-grade encapsulation and transducer architecture.

Integration Notes and Recommended Shortlists

On interfaces, the cleanest system architectures are easy to separate. If your receiver DSP or MCU already has PDM microphone inputs, choose IM69D130 or MP23DB02MM. If it prefers I²S, the ICS-43434 is

especially attractive because it can connect directly to a processor without a separate audio codec. If you want multi-mic beamforming, choose IM69D130 or ICS-52000, because both explicitly support array- oriented behavior. If you want the lowest-noise analog front end, IM73A135V01 is the strongest analog pick in this review. On the output side, use a receiver amplifier for Knowles BA parts, but assume a dedicated piezo/solid-state driver for xMEMS or USound parts.

If the device must both play remote speech and also listen locally, plan for a receiver DSP path that includes at least packet buffering, G.729 decode, AGC/limiting, EQ, and echo management. For noisy environments, dual matched mics are worth the added complexity. Infineon explicitly publishes phase and sensitivity matching for IM69D130 and tight tolerance for IM73A135V01, which is exactly what helps beamforming and stable calibration.

Recommended Shortlist of Complete Mic and Receiver Pairs

Pair

Why it is a strong match

Main trade-offs

Best use case

Sources

Infineon IM73A135V01 +

Knowles

RLQ-34240-000

Highest-confidence speech-first combination in this review. The mic has 73 dB(A) SNR, very low current, and IP57 robustness; the RLQ gives very high output in a

still-small full-range BA package with simpler low- voltage drive than piezo MEMS.

Requires an analog mic front end and a conventional audio amplifier stage; output device is BA, not MEMS.

Best overall if your real requirement is intelligible remote voice in a tiny ear-worn unit.

Infineon IM69D130 +

xMEMS Cowell

Best “digital in, solid- state out” pair. IM69D130 is excellent for DSP-heavy systems and matched arrays; Cowell is the smallest full-range xMEMS speaker publicly aimed at hearing aids and IEMs.

Needs a dedicated Aptos2 piezo amplifier path; public numeric SPL data are less transparent on the current Cowell page than on BA or USound parts.

Best if you explicitly want a MEMS speaker rather than a BA receiver.

TDK ICS-43434 +

USound Achelous UT-P 2018

Cleanest path when you want I²S microphone integration and a MEMS speaker with unusually complete public output and power data. Achelous publishes coupler SPL, capacitance, drive voltage, and power.

Larger receive transducer than Cowell/RAQ; requires 15 Vp class piezo drive and bias handling.

Best documented MEMS-to-MEMS-

ish path for a custom proof-of- concept.

If I had to recommend one path for your exact use case without any extra constraints, it would be Pair 1: IM73A135V01 + RLQ-34240-000. It is the strongest balance of low capture noise, compactness, practical drive electronics, and speech-oriented output. If you insist that the ear output must be a MEMS speaker chip rather than a balanced-armature receiver, then Pair 2 is the best fit. If you specifically want a design built around publicly published MEMS-speaker electrical numbers, Pair 3 is the best-documented route.

The final engineering takeaway is this: ear-canal and hearing-aid-style packaging is feasible with today’s commercial parts; true implantation is not realistically feasible with these off-the-shelf packages; and for G.729 voice, a strong tiny BA receiver still beats most MEMS-speaker paths for simplicity and speech-first practicality.