Voice Recognition Module V3

User manual | English edition | R4

Train short voice commands, load the commands you need, and receive a record ID when a trained command is recognized. VR3 performs recognition on the module and is designed for speaker-dependent operation.

80 stored records	7 active commands	1.5 s per command
Record IDs 0-79 (decimal).	Choose up to seven trained records for the active recognizer.	Use a short word or phrase and record it twice during training.

For an Arduino application, use the examples on pages 9-10. For another host controller, use the UART protocol reference on pages 11-24.

Read this first

A map of the manual and the three operations every application needs.

Figure 1. Training, loading and recognition are separate operations.

Section	Pages	Use it when you need to...
Product and concepts	3-5	Check electrical requirements and understand records, slots and serial routes.
Direct PC testing	6-8	Connect a USB-TTL adapter, train two commands and read recognition events.
Arduino examples	9-10	Compile the supplied examples and use their text commands.
UART protocol	11-20	Build a host driver or inspect complete command frames.
Groups and events	21-24	Use external group selection, user groups and event decoding.
Troubleshooting and reference	25-26	Check common failures, test your application and consult the revision history.

Keep these terms separate

Record: a voice template stored in Flash. Slot: one of seven positions in the active recognizer. Signature: a text label associated with a record. Group: a set of record IDs used to replace the active list.

Product specifications

Electrical, mechanical and recognition specifications.

Item	Specification
Supply voltage	4.5-5.5 V
Current	Less than 40 mA
Digital interface	5 V TTL UART and GPIO
Microphone interface	3.5 mm mono microphone connector and microphone pin interface
Board size	31 mm x 50 mm
Stored voice records	Up to 80; record IDs 0-79 (decimal)
Simultaneously active records	Up to 7
Maximum command duration	1500 ms; typically a short word or phrase
Recognition accuracy	99% under ideal conditions

The current and accuracy specifications do not include detailed measurement conditions. Actual results depend on the operating environment, speaker and microphone placement.

Interfaces and labels

Label	Function
VCC / GND	Module supply and common ground.
TXD / RXD	UART output from VR3 / UART input to VR3. Cross these to the host receiver/transmitter.
IN0-IN2	External group-selection inputs.
OUT0-OUT6	Recognition-related digital outputs; confirm slot-to-output mapping before integration.

Use the labels printed on your board. Board photographs are for orientation; they are not a dimensioned footprint or a connector pin-number specification.

Records and the active recognizer

Store up to 80 templates; activate up to 7 of them at a time.

Figure 2. Stored records and active slots serve different purposes.

Train a record to store a voice template. Load a trained record to make it eligible for recognition. A trained record that has not been loaded is not part of the current recognition set. Sounds can also be trained; short, distinct words are easier to use consistently.

Operation	Stored records	Active recognizer
Train record 7	Creates or updates record 7.	Training alone does not define your final active list.
Load 0, 1, 7	Templates remain stored.	Makes those trained records active.
Clear recognizer	Saved templates remain available.	Empties the list and stops recognition.
Power cycle	Check stored training status after restart.	Reload, or use configured power-on auto load.

Groups let you select a different set of records for the same seven-slot recognizer. Loading a group replaces the active list; it does not add another recognition engine.

Choose your serial route

The PC-facing speed and the module-facing speed are different in the Arduino examples.

Figure 3. The Arduino text-command route and the direct PC route.

Route	PC sends	Serial setup
Arduino training example	Text such as sigtrain 0 ON. The sketch constructs the module frame.	PC to Arduino: 115200, newline enabled. Arduino to VR3: 9600 by default.
Direct USB-TTL	Complete binary frames, entered in HEX-send mode.	PC to VR3: 9600 by default; 8 data bits, no parity, 1 stop bit, no flow control.
Arduino bridge example	Text containing command/data HEX bytes only. The sketch adds the framing.	PC to Arduino: 115200, newline enabled.

Direct PC test: connect and check

Use this route to test VR3 without an Arduino board.

Figure 4. Four-wire USB-TTL connection. The adapter must be compatible with 5 V TTL.

1. Disconnect USB power. Connect 5 V to VCC, GND to GND, adapter TXD to VR3 RXD, and adapter RXD to VR3 TXD. Plug the microphone into VR3, then connect USB to the PC.

2. Install the adapter driver if needed and identify its serial port. Open a serial utility that supports raw HEX transmission, such as AccessPort.

Setting	Value
Baud rate	9600 bit/s by default
Data / parity / stop	8 / None / 1
Send	HEX; no automatically appended CR or LF; periodic send off
Receive	HEX for validation; character view for training prompts

TX | Query system settings

AA 02 00 0A

A settings reply starts with AA 08 00 and ends with 0A. Other bytes depend on the saved settings. If there is no reply, check wiring, the port, the signal level and the baud rate before training.

Direct PC test: train two commands

Train one record at a time and follow the module's microphone prompts.

Record	Spoken command	Signature bytes
0	On	ON = 4F 4E
1	Off	OFF = 4F 46 46

TX | Train record 0 and assign signature ON

AA 05 21 00 4F 4E 0A

Figure 5. Training prompts and LED indications.

When Speak now appears, say "On" once. When Speak again appears, repeat the same command. Use the module microphone, keep a stable voice and wait for the recording indication. If the two recordings do not match, follow the next prompt or retry after timeout.

TX | Train record 1 and assign signature OFF

AA 06 21 01 4F 46 46 0A

Repeat the process, saying "Off" both times. The signature is only a label; it does not determine what you must say. Each spoken command must fit within 1.5 seconds. Both LEDs flash rapidly on failure and more slowly after successful training.

Direct PC test: load and recognize

A successful training operation is followed by loading the trained records.

RX | Successful training examples for records 0 and 1

AA 07 21 01 00 00 4F 4E 0A
AA 08 21 01 01 00 4F 46 46 0A

After 21: N is the successful-training count, followed by the record ID, its result and the signature. For one successful record, N=01 and the result is 00.

TX | Query the training status of records 0 and 1

AA 04 02 00 01 0A

In a 02 reply, inspect each record/status pair: status 01 means trained and 00 means untrained. Do not confuse that status with training-result 00, which means success. See page 14 for the full reply layout.

TX | Clear the active list, then load 0 and 1 (send separately)

AA 02 31 0A
AA 04 30 00 01 0A

RX | Clear acknowledgement and two-record load result

AA 03 31 00 0A
AA 07 30 02 00 00 01 00 0A

Loading starts recognition immediately. Say "On" or "Off". The module reports the matched record without another start command. Keep the port open and view incoming bytes in HEX.

RX | Recognition examples, outside group mode

AA 09 0D 00 FF 00 00 02 4F 4E 0A
AA 0A 0D 00 FF 01 01 03 4F 46 46 0A

In these examples, the record field identifies 0 or 1; the final bytes encode ON or OFF. Unmatched speech may produce no recognition frame.

Arduino: hardware and build

The supplied examples use an UNO and SoftwareSerial on digital pins 2 and 3.

Figure 6. UNO wiring for VR myVR(2,3): D2 is receive, D3 is transmit.

Download the official VoiceRecognitionV3 library. In the following example, its folder is named VR3lib. Copy the vr_sample_train example folder beside it. Install Arduino CLI and the Arduino AVR core for an UNO; another board needs its own core and FQBN. A Sensor Shield V07 is optional; direct four-wire connection is sufficient.

CLI | Install the UNO platform, if it is not already installed

arduino-cli core update-index
arduino-cli core install arduino:avr

CLI | Compile the copied training example

arduino-cli compile -b arduino:avr:uno --library VR3lib vr_sample_train

CLI | Upload to the UNO; replace COM5 with the actual board port

arduino-cli upload -b arduino:avr:uno -p COM5 vr_sample_train

Open the UNO serial port at 115200 bit/s in a terminal that can append a newline. The example initializes the module-facing port at 9600 bit/s. The UNO and VR3 share ground and 5 V power as shown above.

Arduino: train and use the examples

Text commands on this page are entered through the Arduino training sketch.

TEXT | Send one line at a time; PC port at 115200

settings
sigtrain 0 ON
sigtrain 1 OFF
load 0 1

For each sigtrain command, wait for Speak now and Speak again and say the intended command twice. Use record 0 for "On" and record 1 for "Off". After load 0 1, say either command and inspect the printed recognition result.

Training-sketch command	Purpose
record 0 1	Check whether records 0 and 1 are trained.
vr	Inspect the active recognizer.
getsig 0	Read the label associated with record 0.
clear	Stop recognition and empty the active list.
help	Print the sketch's supported command list.

Example selection

Example	What it demonstrates
vr_sample_control_led	Uses trained records 0 and 1 to control the UNO LED. Prepare both records before running it.
vr_sample_bridge	Adds framing around HEX command/data text. For settings send 00; for clear send 31; for loading 0/1 send 30 00 01.
vr_sample_multi_cmd	Uses record 0 as an application-level switch between 0-6 and 0,7-12. Only one set is active at a time.
vr_sample_check_baud_rate	Helps find a previously changed module baud rate.

UART frame format

All protocol numbers are HEX bytes unless a value is explicitly marked decimal.

Figure 7. A complete load frame and the length calculation.

Field	Meaning
AA	Frame head.
LEN	Number of bytes in LEN + command + data.
Command	One command or event byte.
Data	Zero or more command-specific bytes.
0A	Frame terminator at the position determined by LEN.

RULE | Let p be the number of data bytes

LEN = 2 + p
TOTAL_BYTES = LEN + 2

There is no checksum field in the documented frame. HEX-send mode converts each written pair, such as AA, into one byte. Sending the characters "AA" in ordinary text mode sends different bytes.

Receive by length, not by line endings

Find AA, read LEN, then collect LEN more bytes after the first two bytes. Verify that the last byte is 0A and that the command/data layout is valid. A serial read may contain part of a frame, one frame or several frames.

Command map and notation

Use one set of count-based symbols throughout the protocol reference.

Code	Command family	Purpose
00-03	Queries	Settings, recognizer, training status and signature.
10-15	System settings	Restore, baud rate, output mode, pulse width, output reset and auto load.
20-22	Record operations	Train, train with signature and edit signature.
30-32	Recognizer and groups	Load, clear and group operations.
0A / 0D / FF	Module-originated messages	Training prompt, recognition event and protocol error.

Symbol	Definition
m	Number of record IDs or record/status pairs in this frame; m is a count, not a last index.
k	Number of signature bytes.
t	Number of prompt bytes, including any CR byte included by the module.
R0...R(m-1)	Record IDs. Use 00-4F for the 80-record product range.
N	Count/status header described by the relevant command. Do not reuse one command's status meaning for another.
[expression]	The numeric value must be calculated and encoded as one HEX length byte. Brackets are not transmitted.

For each command, count the data fields shown in the layout and calculate LEN before sending the frame.

Queries: settings and recognizer

00 reports settings; 01 reports the current active list.

00 | TX and symbolic RX

AA 02 00 0A
AA 08 00 ST BR IOM PW AL GRP 0A

Field	Interpretation
ST	00: successful settings-query response; not a record-training status.
BR	00 or 03: 9600; 01: 2400; 02: 4800; 04: 19200; 05: 38400 bit/s.
IOM / PW	Output mode / pulse-width level. IOM: 00=Pulse, 01=Toggle, 02=Down/Clear, 03=Up/Set.
AL	Saved auto-load bitmap: 00 disables it; 01/03/07/.../7F select active list positions. Erased FF is disabled at boot.
GRP	Saved external control mode: 00=off, 01=system, 02=user. Restart to apply a saved change.

01 | TX and symbolic RX

AA 02 01 0A
AA 0D 01 RVN R0 R1 R2 R3 R4 R5 R6 RTN MAP GRPM 0A

Field	Interpretation
RVN	Number of valid active records, at most 7.
R0-R6	Record ID occupying each slot; FF denotes an unused/unset slot.
RTN / MAP	Total listed records / valid-slot bitmap. Bit i of MAP corresponds to slot i.
GRPM	FF: non-group mode; 00-0A: system-group range; 80-87: user-group indication.

Queries: training status and signature

02 checks stored voice templates; 03 reads a record's text label.

02 | TX: specified records, or all records

AA [2+m] 02 R0 ... R(m-1) 0A
AA 03 02 FF 0A

02 | RX: m record/status pairs in the returned frame

AA [3+2*m] 02 N R0 STA0 ... R(m-1) STA(m-1) 0A

Status in a 02 record/status pair	Meaning
00	Untrained
01	Trained
FF	Record ID outside the supported range

For a specified-record query, N counts trained records in that request. An all-record query returns five pairs per frame and N is cumulative across frames. To check the 80-record product range, query IDs 0-79 explicitly in small batches, such as five records.

02 | TX: check records 0 and 1

AA 04 02 00 01 0A

03 | TX and symbolic RX

AA 03 03 R 0A
AA [4+k] 03 R k SIGNATURE 0A

Here k is the number of signature bytes. The reply includes the record byte and signature-length byte, so LEN=4+k.

03 | Example: query record 0; reply contains signature ON

AA 03 03 00 0A
AA 06 03 00 02 4F 4E 0A

Command 21 stores at most 10 signature bytes and reports F0 on truncation. Command 22 accepts 0-9 bytes. Use short ASCII labels for predictable byte counts.

System settings: restore and baud rate

10 restores settings; 11 changes the module UART rate after restart.

10 | Restore system settings; acknowledgement

AA 02 10 0A
AA 03 10 00 0A

11 | Set baud rate; acknowledgement

AA 03 11 BR 0A
AA 03 11 00 0A

BR (HEX)	Baud rate (bit/s)
00 or 03	9600
01	2400
02	4800
04	19200
05	38400

11 | Example: select 19200 for the next restart

AA 03 11 04 0A

Receive the acknowledgement at the current baud rate. Restart VR3, then change the host to the selected rate and query settings again. If using an Arduino sketch, update the module-facing SoftwareSerial rate as well.

If the saved rate is unknown, try the supported rates using the read-only settings query. Change only the host rate while searching; do not send blind configuration commands.

GPIO configuration commands

12 selects output behavior; 13 selects pulse duration; 14 resets outputs.

12 | Set output mode; acknowledgement

AA 03 12 MODE 0A
AA 03 12 00 0A

MODE	Output behavior
00	Pulse: active-low pulse on recognition.
01	Flip / Toggle: alternate output state on each event.
02	Down / Clear: drive low on recognition; retain until reset.
03	Up / Set: drive high on recognition; retain until reset.

13 / 14 | Symbolic commands and acknowledgements

AA 03 13 LEVEL 0A
AA 03 13 00 0A
AA [2+m] 14 IO0 ... IO(m-1) 0A
AA 03 14 00 0A

LEVEL -> width	LEVEL -> width	LEVEL -> width	LEVEL -> width
00 -> 10 ms	04 -> 30 ms	08 -> 50 ms	0C -> 300 ms
01 -> 15 ms	05 -> 35 ms	09 -> 75 ms	0D -> 400 ms
02 -> 20 ms	06 -> 40 ms	0A -> 100 ms	0E -> 500 ms
03 -> 25 ms	07 -> 45 ms	0B -> 200 ms	0F -> 1000 ms

For 14, use output IDs 00-06, or a single FF to reset all outputs. Pulse-width levels range from 00 through 0F.

Understanding the output waveforms

Mode 02 is Down/Clear; mode 03 is Up/Set.

Figure 8. Output states during recognition and reset; timing is not to scale.

Marker / mode	Meaning
A	Records are loaded; initial output state is shown schematically.
B / C / D	The associated recognition event occurs.
E	Host sends output reset.
Pulse	The output goes low briefly and returns high after the configured duration.
Toggle	Each event changes high to low or low to high.
Down / Clear	Recognition makes the output low; reset returns it to its initial high state.
Up / Set	Recognition makes the output high; reset returns it to its initial low state.

14 | Examples: reset all outputs / reset IO1 (send separately)

AA 03 14 FF 0A
AA 03 14 01 0A

Power-on auto load

15 stores the record list to activate after the module starts.

15 | Configure m records; symbolic acknowledgement

AA [3+m] 15 BITMAP R0 ... R(m-1) 0A
AA [4+m] 15 00 BITMAP R0 ... R(m-1) 0A

Number of entries m	BITMAP (HEX)
0 / disabled	00
1	01
2	03
3	07
4	0F
5	1F
6	3F
7	7F

For a contiguous list of m entries, the bitmap is (1 << m) - 1. It describes list positions, not the record-ID value. Train the records first, then configure auto load.

15 | Example: auto-load records 0 and 1; acknowledgement

AA 05 15 03 00 01 0A
AA 06 15 00 03 00 01 0A

15 | Disable auto load; acknowledgement

AA 03 15 00 0A
AA 04 15 00 00 0A

Record commands: train and label

20 records voice templates; 21 also assigns a signature; 22 edits a signature.

20 | Train m records; final result with m record/status pairs

AA [2+m] 20 R0 ... R(m-1) 0A
AA [3+2*m] 20 N R0 STA0 ... R(m-1) STA(m-1) 0A

During training, the module sends 0A prompt frames. N in the final result is described as the successful-training count. Use a one-record operation first so that prompt and result association is unambiguous.

21 | Train record R with k signature bytes; final result

AA [3+k] 21 R SIGNATURE 0A
AA [5+k] 21 N R STA SIGNATURE 0A

22 | Set signature / delete signature (symbolic TX)

AA [3+k] 22 R SIGNATURE 0A
AA 03 22 R 0A

22 | Signature-set / signature-delete symbolic RX

AA [4+k] 22 00 R SIGNATURE 0A
AA 04 22 00 R 0A

Final training result STA	Meaning
00	Trained successfully.
F0	Command 21 trained successfully; signature truncated to 10 bytes.
FE	Training failed after the retry limit.
FF	Invalid record value.

Recognizer commands: load and clear

30 activates trained records; 31 removes the current recognition list.

30 | Load m records; reply with m result pairs

AA [2+m] 30 R0 ... R(m-1) 0A
AA [3+2*m] 30 N R0 STA0 ... R(m-1) STA(m-1) 0A

Load result	Meaning
00	Loaded successfully
FC	Record already in the active recognizer
FD	Recognizer full
FE	Record is not trained
FF	Record ID outside the supported range

N is the number loaded successfully. A load operation begins recognition immediately. The active recognizer can hold at most seven valid records. If the old list is no longer wanted, clear it before loading a replacement.

30 | Example: load records 0, 1 and 2; all successful

AA 05 30 00 01 02 0A
AA 09 30 03 00 00 01 00 02 00 0A

31 | Clear active recognizer; acknowledgement

AA 02 31 0A
AA 03 31 00 0A

Clear stops recognition and empties the active list. It does not erase trained templates. The yellow LED flashes slowly during recognition and stays on after the recognizer is cleared.

Groups and external selection

A group is a record list, not another recognition engine.

Figure 9. Fixed system groups versus configurable user groups.

System groups contain fixed consecutive ranges; user groups contain custom lists of up to seven records. Loading either group replaces the active recognition list. User groups are numbered 0-7.

IN2	IN1	IN0	Selected group (decimal)
LOW	LOW	LOW	7
LOW	LOW	HIGH	6
LOW	HIGH	LOW	5
LOW	HIGH	HIGH	4
HIGH	LOW	LOW	3
HIGH	LOW	HIGH	2
HIGH	HIGH	LOW	1
HIGH	HIGH	HIGH	0

The product supports 80 records, IDs 0-79. System groups 00-0A cover complete seven-record sets through record 76. Load records 77-79 directly by record ID.

Group commands: configure

32 uses a subcommand byte; lengths include that byte.

32 / 00 | Set external group mode; query current mode

AA 04 32 00 MODE 0A
AA 04 32 00 FF 0A

MODE	Saved configuration
00	Disable external group control
01	System-group selection by inputs
02	User-group selection by inputs; see the integration note below
FF	Query only

32 / 00 | Set acknowledgement / query reply

AA 03 32 00 0A
AA 05 32 00 FF MODE 0A

The command saves the mode to Flash; restart to apply it. Verify external pin-based user-group selection on the target module before integration. For the initial user-group test, use UART command 32/03.

32 / 01 | Set m entries in a user group; delete that group

AA [4+m] 32 01 GID R0 ... R(m-1) 0A
AA 04 32 01 GID 0A

32 / 01 | Example: group 0 contains records 0, 7 and 25

AA 07 32 01 00 00 07 19 0A

25 is decimal here; its HEX record byte is 19. A successful configuration is acknowledged with AA 03 32 00 0A. Editing a group describes its contents; use a load-group command to activate it.

Group commands: load and inspect

Use a loaded-list response to verify which records became active.

32 / 02 and 03 | Load a system group / a user group

AA 04 32 02 GID 0A
AA 04 32 03 GID 0A

32 | Group-load reply layout

AA 0D 32 GID R0 R1 R2 R3 R4 R5 R6 RTN MAP GRPM 0A

GID is the loaded group ID; R0-R6 are the listed records; RTN, MAP and GRPM describe the list. Both system-group and user-group load replies use this layout and LEN=0D.

32 / 04 | Query all user groups / one user group

AA 03 32 04 0A
AA 04 32 04 GID 0A

32 / 04 | One group-content reply

AA 0A 32 GID R0 R1 R2 R3 R4 R5 R6 0A

A query for all groups can return several frames. Associate each frame with its group ID. FF denotes an unset group entry; do not use an untrained or unset record as a recognition template.

32 | Examples: query user group 0, then load user group 0

AA 04 32 04 00 0A
AA 04 32 03 00 0A

Module-originated messages

Training prompts, recognition events and protocol errors can arrive asynchronously.

0A | Training prompt, with t prompt bytes

AA [3+t] 0A R PROMPT 0A

Prompt text includes Speak now, Speak again and Success. A prompt may include a CR byte before the final terminator. Read the length and treat the prompt as bytes. Its length depends on the text; the command byte 0A is separate from the final frame terminator.

0D | Recognition event, with k signature bytes

AA [7+k] 0D 00 GRPM R RI k SIGNATURE 0A

Figure 10. Record 0 recognized with the ASCII signature ON.

GRPM is FF outside group mode. R is the trained record ID, RI is its active slot and k is the signature length. If k=0, no signature bytes follow. The fixed byte after command 0D is 00.

FF | Error response

AA 03 FF ECODE 0A

ECODE	Meaning
FF	Undefined command
FE	Command length error
FD	Invalid data
FC	Invalid subcommand
FB	Command usage error

Log the complete frame when an error occurs. Error-frame ECODE and training-result STA belong to different commands; the same byte value need not have the same meaning.

Troubleshooting and validation

Work from electrical checks to communication, training, loading and recognition.

Symptom	What to check
No serial port	USB data cable, adapter connection and driver installation. Identify the actual port rather than assuming COM5.
No settings reply	Power, common ground, crossed TX/RX, compatible TTL level, correct port, HEX mode and complete frame.
Unreadable receive text	Protocol bytes are binary. Use HEX view; only the prompt/signature portions are readable text. Verify baud rate.
Training fails	Microphone plugged into VR3, correct timing, quiet environment, stable speech and a command shorter than 1.5 s.
Trained but not recognized	Query training status, inspect the active list and the load results. Test using the person who trained the record.
Wrong command / false trigger	Use more distinct words; retrain at a representative speaking distance and noise level.
Works until power is removed	Verify saved training status, baud rate and auto-load settings after restart. Reload records if required.
Unexpected GPIO behavior	Check the selected mode (02=Down/Clear, 03=Up/Set), output/slot mapping, wiring and reset sequence.

A practical acceptance check

Test both commands at least ten times each and record correct matches, missed matches and wrong matches. Repeat in the intended environment. Use the results to assess your application; this check does not establish the ideal-condition accuracy specification.

Also test power cycling, a deliberately untrained record, an already-loaded record, switching between record lists, and malformed-frame recovery in your own host driver. Save raw HEX logs together with the hardware/firmware identification.

Revision History

Document revisions describe changes to this manual, not a hardware or firmware version.

Rev.	Date	Changes
R4	15 Oct 2026	Moved change descriptions to this final chapter. Rewrote the main text as current operating instructions. Kept source-review details in the separate Chinese review. No change to R3 protocol definitions or the 80-record capacity.
R3	15 Oct 2026	Aligned the protocol with the confirmed production firmware: GPIO 02=Down/Clear and 03=Up/Set; final training results FE/FF and signature-truncation result F0; active-low group inputs; clarified settings, counts and output-slot mapping. Capacity remains 80 records.
R2	15 Oct 2026	Interim clarification of product capacity and legacy protocol descriptions. Protocol assignments in this edition were superseded by R3.
R1	15 Oct 2026	Rewrote the English manual, separated the direct PC and Arduino routes, added ten explanatory diagrams, and standardized terminology and frame-length notation.