What is audio intelligence?

Audio intelligence is turning recordings into structured data a system can act on: who spoke, what was said, what was agreed, and what should happen next. With bem you define the record you want as a JSON schema, send the recording, and get that record back with an audit trail on the run.

What kinds of audio can bem read?

Any recording a person could listen to and take notes from. The examples on this page are an earnings call, an insurance claim call, a medical consultation, an air traffic control tape, and calls to a bank's phone line. The same approach covers voicemail, dictation, hearings, public meetings, and radio.

Which audio formats and file sizes does bem accept?

MP3, WAV, and M4A. A workflow call accepts files up to 50 MB each, sent as a multipart upload or as base64 JSON. The audio track of an MP4 or MOV video is read the same way.

Does bem transcribe the recording?

Yes. Speech is transcribed before extraction, and the record is filled in from what was said. If you want the transcript itself, ask for it in your schema.

How is this different from a transcription service?

A transcription service returns the words. bem returns the record: the figures from an earnings call as numbers, a claim as a first notice of loss, a call scored against your checks. The words come with it, as the evidence for each field.

Can bem score calls against our own quality scorecard?

Yes. Each line of your scorecard is a field in the schema, with a description of what counts as meeting it. Ask for the agent's words as well and every line that is met comes back with its quote. One of the calls on this page is scored that way.

How does it handle a queue with many kinds of recording?

A Classify function sorts each recording by what it is, and sends it to the Extract function for that kind. The queue on this page is 15 calls of 7 kinds, filed by one workflow.

How long does it take to process a recording?

In the recorded runs on this page, a single Extract function took between 3.7 and 12.7 seconds on recordings of 27 seconds to two and a half minutes. Each example shows the time its run took.

Per function call, the same as any other file. There is no per-minute meter, and the first 100 function calls are free.

Use cases · Audio

Audio intelligence for critical industries.

Earnings calls, claim calls, consultations, radio, and phone lines go in. Typed records come out, each field with the words it rests on.

Start building free Talk to an engineer

bem returns6.1 s

company: Alphabet
quarter: Q3
revenue_growth_yoy_pct: 25
annual_revenue_run_rate: half a trillion dollar
search_query_growth: highest it's been in the last five years
google_cloud_revenue_growth_pct: 103
cloud_growth_driver: AI infrastructure and solutions
capacity_plan: bring more capacity online to meet this long-term demand

0:54 / 0:548 of 8 fields

Alphabet Q3 2026 earnings call, Sundar Pichai, 0:54 (Alphabet Investor Relations). The record is a bem run on it. bem is not affiliated with or endorsed by Alphabet.

01Critical industries

Earnings calls, claims, consultations, and radio, read as they play.

Four recordings, four kinds of record. Each tape runs as you reach it, with the sound off until you ask for it. The words a field rests on turn blue as they are said, and the field's row lights up. Press a row to hear the words behind it.

01Earnings callInvestment research, investor relations

Earnings call, as broadcast

0:00 / 1:28

A CFO reads out the quarter. Eleven figures come back as fields.

Alphabet's CFO goes through the third quarter in under a minute and a half: revenue, costs, expenses, operating income, and margin, each with its change on the year. bem returns every figure as a number in its own field, with a decrease as a negative.

The schema asks for

quarters_of_double_digit_growth
consolidated_revenue_usd_bn
revenue_growth_pct
cost_of_revenue_usd_bn
traffic_acquisition_costs_usd_bn
operating_expenses_usd_bn
rd_expense_growth_pct
ga_expense_change_pct
operating_income_usd_bn
operating_income_growth_pct
operating_margin_pct
the words each one rests on

What it feedsA model that is updated when the call ends, and a record of what management said to set against the next quarter.

Alphabet's Q3 2026 earnings call, the CFO's results: 1:28 of a call of an hour, from Alphabet's own upload. The tape prints the words bem gave as evidence. bem is not affiliated with or endorsed by Alphabet.Alphabet Investor Relations

bem returns6.8 s

field	value	the words it rests on
quarters_of_double_digit_growth	13	“13th consecutive quarter”
consolidated_revenue_usd_bn	127.6	“$127.6 billion”
revenue_growth_pct	25	“25%”
cost_of_revenue_usd_bn	51.4	“$51.4 billion”
traffic_acquisition_costs_usd_bn	16.4	“$16.4 billion”
operating_expenses_usd_bn	35.6	“$35.6 billion”
rd_expense_growth_pct	36	“36%”
ga_expense_change_pct	-17	“decreased 17%”
operating_income_usd_bn	40.6	“$40.6 billion”
operating_income_growth_pct	30	“30%”
operating_margin_pct	31.8

Recorded bem run, one Extract function. Not every field is shown. The words on the tape are the ones bem gave as evidence.

02Claim intakeInsurance carriers, claims administrators

Generated call

0:00 / 1:21

A roadside call becomes the first notice of loss.

A driver reports being hit from behind, twenty minutes after it happened. In 81 seconds the call covers the policy, the place, the other driver, the police report, and the damage. bem returns the claim record, with the words behind every field.

The schema asks for

policy_number
policyholder
loss_type
time_of_loss
location
anyone_injured
other_driver
other_vehicle
other_plate
other_insurer
police_report_number
vehicle_drivable
damage
claim_number
next_step
the words each one rests on

What it feedsThe claim file, open before the caller hangs up: the coverage check, the repair booking, and the first fraud screen.

No real claim call is public, so this one was written and voiced for this page (text-to-speech voices, on a telephone band) and can be played in full. The record is what bem returned for it.

bem returns8.0 s

field	value	the words it rests on
policy_number	NM-447-2190	“NM-447-2190”
policyholder	Mr. Okafor	“Mr. Okafor”
loss_type	rear-ended	“rear-ended”
time_of_loss	5:40	“540”
location	Route 9 at the light by Maple Avenue	“Route 9 at the light by Maple Avenue”
anyone_injured	false	“we're both okay”
other_driver	Linda Marsh	“Linda Marsh”
other_vehicle	gray pickup	“gray pickup”
other_plate	KTR-8814	“KTR-8814”
other_insurer	Kestrel Insurance	“Kestrel Insurance”
police_report_number	26-11479	“26-11479”
vehicle_drivable	true	“it drives”
damage	The trunk won't close, and a taillight is out	“The trunk won't close, and a taillight is out”
claim_number	CL-773-052	“CL-773-052”
next_step	An adjuster will call you within one business day	“An adjuster will call you within one business day”

Recorded bem run, one Extract function. Every field is shown. The words on the tape are the transcript bem returned with it.

03ConsultationProviders, telehealth, clinical documentation

Mock consultation, pauses shortened

0:00 / 2:20

A history told out of order becomes the fields a clinician reads first.

A video consultation about lips that swelled after a meal. The patient is unsure of the times and of what she takes. bem returns what was eaten and when, how fast the swelling came on, whether her breathing is affected, and her medicines and allergies.

The schema asks for

suspected_trigger
time_eaten
time_to_swelling
swelling_getting_worse
difficulty_breathing
chest_tightness
history_of_asthma
current_medication
other_inhalers
antihistamine_use
known_allergies
the words each one rests on

What it feedsThe draft note for the clinician to check and sign, and the medicine and allergy lists on the record.

Two passages of a mock primary-care consultation of 8:30, with a practising clinician and an actor as the patient, from the PriMock57 corpus by Babylon Health. Long pauses were shortened. The tape prints the words bem gave as evidence.PriMock57, CC BY 4.0

bem returns11.4 s

field	value	the words it rests on
suspected_trigger	prawn soup	“prawn soup”
time_eaten	around seven	“around seven”
time_to_swelling	half an hour	“half an hour”
swelling_getting_worse	true	“getting worse”
difficulty_breathing	true	“can't breathe that well”
chest_tightness	unsure	“I'm not sure”
history_of_asthma	true	“generally have asthma”
current_medication	Salbutamol inhalers	“Salbutamol inhalers”
other_inhalers	no	“No”
antihistamine_use	not sure which ones	“I don't know which ones”
known_allergies	peanut allergy	“peanut allergy”

Recorded bem run, one Extract function. Not every field is shown. The words on the tape are the ones bem gave as evidence.

04Tower tapeAviation safety, incident review

FAA recording, as published

0:00 / 2:28

Two and a half minutes of radio become the incident record.

New York approach control, 15 January 2009. US Airways 1549 reports a bird strike after take-off, is offered runways at two airports, and comes down on the Hudson. Everyone on board survived. The radio is narrow, clipped, and shared with other aircraft. bem returns what was offered and what the crew answered.

The schema asks for

nature_of_emergency
engines_affected
first_intention
laguardia_runway_31_response
laguardia_runway_4_response
alternate_airport_requested
teterboro_runway_offered
teterboro_response
final_stated_intention
radar_contact_lost
the words each one rests on

What it feedsThe review file: each option offered and each answer, ready to set against the radar track and the flight data.

The first 2:28 of the FAA's recording of New York TRACON, as the FAA published it. It is in the public domain. The tape prints the words bem gave as evidence.FAA accident and incident data

bem returns10.9 s

field	value	the words it rests on
nature_of_emergency	bird strike	“bird strike”
engines_affected	both engines	“both engines”
first_intention	returning back towards LaGuardia	“returning back towards LaGuardia”
laguardia_runway_31_response	Unable.	“Unable.”
laguardia_runway_4_response	I'm not sure we can make any runway	“I'm not sure we can make any runway”
alternate_airport_requested	Teterboro	“Teterboro”
teterboro_runway_offered	runway 1	“runway 1”
teterboro_response	We can't do it.	“We can't do it.”
final_stated_intention	We're going to be in the Hudson.	“We're going to be in the Hudson.”
radar_contact_lost	true	“radar contact is lost”

Recorded bem run, one Extract function. Not every field is shown. The words on the tape are the ones bem gave as evidence.

02Contact centers

Phone lines: the wrap-up form and the quality check, from the call.

Three real calls to a bank's phone line. The first becomes the form the agent would have typed. The other two are checked against what the agent should have done, and one of them is the kind a sample misses: it sounds fine, and the booking is wrong.

05After-call workRetail banking, credit unions

Real call, pauses shortened

0:00 / 1:19

A transfer request becomes the form the agent would have typed.

The caller wants $122 moved from checking to savings. The agent takes the amount and the two accounts, then explains when the money will arrive and how to speed it up. bem returns the form, with the words each field rests on.

The schema asks for

caller_name
agent_name
source_account
destination_account
amount_usd
outcome
the words each one rests on

What it feedsThe transfer ticket and the note on the customer's record, with nothing left for the agent to type once the call ends.

A call to a bank's phone line, 1:54 as recorded, from the HarperValleyBank corpus by Gridspace and Stanford. Long pauses were shortened.HarperValleyBank corpus, CC BY 4.0

bem returns6.8 s

field	value	the words it rests on
caller_name	David Johnson	“David Johnson”
agent_name	Michael	“Michael”
source_account	checking account	“checking account”
destination_account	savings account	“savings account”
amount_usd	122	“$122”
outcome	transferred $122 from checking to savings	“$122 has been transferred from your checking account to your savings account”

Recorded bem run, one Extract function. Every field is shown. The words on the tape are the transcript bem returned with it.

06Read-back checkQuality and compliance teams

Real call, pauses shortened

0:00 / 0:27

The caller asked for 3:30. The agent booked 2:30.

Nothing in this call sounds wrong. The agent is polite, follows the script, and confirms the appointment. bem is asked for the time the caller wanted and the time the agent read back, and whether they match. They do not. The form the agent filled in for this call says 2:30.

The schema asks for

caller_name
agent_name
appointment_day
time_caller_asked_for
time_agent_confirmed
read_back_matches_request
the words each one rests on

What it feedsA supervisor's queue the same day, before the customer arrives an hour after the slot that was booked.

A call to a bank's phone line, 0:53 as recorded, from the HarperValleyBank corpus by Gridspace and Stanford. Long pauses were shortened.HarperValleyBank corpus, CC BY 4.0

bem returns4.1 s

field	value	the words it rests on
caller_name	Linda Miller	“Linda Miller”
agent_name	David	“David”
appointment_day	Tuesday	“Tuesday”
time_caller_asked_for	3:30 PM	“3.30 p.m.”
time_agent_confirmed	2:30 PM	“2.30 p.m.”
read_back_matches_request	false	“3.30 p.m.”

Recorded bem run, one Extract function. Every field is shown. The words on the tape are the transcript bem returned with it.

07ScorecardQuality teams, outsourced contact centers

Real call, pauses shortened

0:00 / 0:34

A call that skips the greeting, scored against the script.

The agent never says the bank's name or their own, then does everything else right. bem marks those two lines not met, and quotes the agent for the three that are.

The schema asks for

greeting_names_the_bank
agent_gives_own_name
agent_confirms_outcome
agent_offers_more_help
agent_thanks_caller
the words each one rests on

What it feedsThe agent's coaching notes and the client's quality report, built from every call and not a sample.

A call to a bank's phone line, 1:15 as recorded, from the HarperValleyBank corpus by Gridspace and Stanford. Long pauses were shortened.HarperValleyBank corpus, CC BY 4.0

bem returns3.7 s

check	result	the words it rests on
greeting_names_the_bank	not met	nothing in the recording
agent_gives_own_name	not met	nothing in the recording
agent_confirms_outcome	met	“All right, $94 has been transferred from your checking account to your savings account.”
agent_offers_more_help	met	“Is there anything else I can help you with?”
agent_thanks_caller	met	“Thank you for calling. Have a great day.”

Recorded bem run, one Extract function. Every check is shown. The words on the tape are the transcript bem returned with it.

03A mixed queue

15 calls of 7 kinds, sorted and filed by one workflow.

A phone line does not take one kind of call. Classify hears what each caller wants and sends the call to the Extract function for that kind of call, so a bill payment comes back as a payment form and a lost card as a card order. Press play on any of them.

01pay-bill
caller_name
David Miller
payee_company
Fossil Gas
payee_address
930 1st Street Forest Ranch Oregon 60297
amount_usd
92
length 0:41bem took 9.5 s
02replace-card
caller_name
James Williams
card_type
credit
length 0:25bem took 6.8 s
03transfer-money
caller_name
Patricia Miller
source_account
checking account
destination_account
savings account
amount_usd
117
length 0:48bem took 8.5 s
04reset-password
caller_name
John Garcia
send_reset_to
phone
phone_number
715-139-0787
length 0:36bem took 8.5 s
05check-balance
caller_name
Patricia Wilson
account_type
checking
balance_usd
81
length 0:32bem took 6.7 s
06order-checks
caller_name
Patricia Johnson
mailing_address
733 First Street, Forest Ranch, Oregon, 40779
length 0:32bem took 8.9 s
07pay-bill
caller_name
Linda Davis
payee_company
Fossil Gas
payee_address
429 First Street, Forest Ranch, Oregon, 39983
amount_usd
143
length 0:37bem took 7.6 s
08get-branch-hours
caller_name
Elizabeth Jones
branch_hours
9.30 a.m. until 5 p.m.
length 0:23bem took 5.0 s
09transfer-money
caller_name
Mary Smith
source_account
savings account
destination_account
checking account
amount_usd
126
length 0:39bem took 9.3 s
10replace-card
caller_name
Robert Wilson
card_type
credit
length 0:29bem took 7.1 s
11reset-password
caller_name
Elizabeth Rodriguez
send_reset_to
phone
phone_number
222-699-0071
length 0:31bem took 6.9 s
12transfer-money
caller_name
Jennifer Brown
source_account
Checking account
destination_account
Savings account
amount_usd
146
length 0:54bem took 24.9 s
13check-balance
caller_name
Patricia Williams
account_type
checking
balance_usd
105
length 0:28bem took 6.8 s
14pay-bill
caller_name
John Miller
payee_company
Fossil Gas
payee_address
871 Main Street, Harper Valley, California, 81623
amount_usd
103
length 0:48bem took 10.0 s
15replace-card
caller_name
Michael Rodriguez
card_type
credit
length 0:22bem took 6.3 s

Recorded bem runs: one workflow, called once for each call. Classify sorts the call, and the Extract function for that kind of call fills in its form. The time is the whole workflow for that call. Not every field is shown. The calls are from the HarperValleyBank corpus (CC BY 4.0), with long pauses shortened.

04One workflow

From a recording to a decision, in one workflow.

Extract turns the recording into a record. The rest of the platform takes it from there: match it to your data, apply your policies, and send the result where the work happens.

Input

AUDRecordingMP3, WAV, or M4A
CRMYour recordsPolicies, customers, accounts

Components

01ClassifyKnows a claim from an earnings call from a consultation, and sends each the right way.
02ExtractTranscribes the recording and returns the fields in your schema, with the words behind them.
03EnrichMatches the caller and the policy to your records.
04DecideApplies your policies to the record.
05SendDelivers the result to your system.

Decision

Fast-track the claim

No one was hurt, the car drives, and a police report is on file (26-11479). Policy: a collision with no injury and a report number goes to fast-track repair, with no adjuster visit.

The facts are from the recorded run on the claim call above. The policy and the decision are an example.

How the platform works Decide: your policies, applied

05For developers

The audio extraction API: a schema and a file.

Define the record you want once. Then send recordings to the same endpoint you would send a PDF to. The response is your schema, filled in. This is the earnings call above, without the transcript and the evidence it was also asked for.

terminalbem API v3

# 1. Say what you want back
curl -X POST https://api.bem.ai/v3/functions \
  -H "x-api-key: $BEM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "functionName": "quarter-results",
    "type": "extract",
    "outputSchemaName": "QuarterResults",
    "outputSchema": {
      "type": "object",
      "properties": {
        "quarters_of_double_digit_growth": {
          "type": "number",
          "description": "How many consecutive quarters, as stated."
        },
        "consolidated_revenue_usd_bn":      { "type": "number" },
        "revenue_growth_pct":               { "type": "number" },
        "cost_of_revenue_usd_bn":           { "type": "number" },
        "traffic_acquisition_costs_usd_bn": { "type": "number" },
        "operating_expenses_usd_bn":        { "type": "number" },
        "rd_expense_growth_pct":            { "type": "number" },
        "ga_expense_change_pct": {
          "type": "number",
          "description": "Negative for a decrease."
        },
        "operating_income_usd_bn":          { "type": "number" },
        "operating_income_growth_pct":      { "type": "number" },
        "operating_margin_pct":             { "type": "number" }
      }
    }
  }'

# 2. Send the recording
curl -X POST https://api.bem.ai/v3/workflows/quarter-results/call \
  -H "x-api-key: $BEM_API_KEY" \
  -F "wait=true" \
  -F "file=@earnings-call.mp3"

responsetransformedContent

{
  "quarters_of_double_digit_growth": 13,
  "consolidated_revenue_usd_bn": 127.6,
  "revenue_growth_pct": 25,
  "cost_of_revenue_usd_bn": 51.4,
  "traffic_acquisition_costs_usd_bn": 16.4,
  "operating_expenses_usd_bn": 35.6,
  "rd_expense_growth_pct": 36,
  "ga_expense_change_pct": -17,
  "operating_income_usd_bn": 40.6,
  "operating_income_growth_pct": 30,
  "operating_margin_pct": 31.8
}

Read the quickstart

06More to build

More audio use cases, wherever something is already being recorded.

Most recordings are kept for a few months and never heard again. A schema turns them into something a system can act on.

Finance
Investor days and conference talks
Targets, guidance, and the questions analysts asked, set against what was said the quarter before.
Finance
Central bank and regulator briefings
Each statement on rates, outlook, and risk, with the words, minutes after the briefing ends.
Insurance
Recorded statements
Each party's account of the loss, question by question, and the points where two accounts differ.
Healthcare
Dictation and handover
A clinician's spoken note becomes the fields of the record, ready to check and sign.
Legal
Hearings and depositions
Who said what, the exhibits named, and the undertakings given, each tied to its place in the recording.
Government
Public meetings
Motions, votes, and who spoke for how long, from the recording a council already publishes.
Defense
Radio and voice nets
Who reported what, from where, and when: a log an analyst can check against the recording.
Logistics
Check calls and dispatch
A driver's check call becomes a status update on the load: location, time of arrival, and any exception.
Collections
Disclosures and promises to pay
Was the required disclosure read, was the right person on the line, and what was promised, with the words.
Sales
Third-party verification
Each statement the script requires and each yes the customer gave, checked before the enrollment is accepted.
Field service
Voice notes from the job
A technician talks through the visit on the way out. The work order is filled in before they reach the truck.
Operations
Voicemail
Every message sorted by what it asks for, with the callback number and the deadline pulled out.

Video intelligence Finance Insurance Healthcare

07Reliability

Audio data you can rely on.

01
Your schema, not ours
You decide the fields. The same recording returns a wrap-up form for operations and a scorecard for quality.
02
The words behind each field
Ask for the quote and every value comes back with the words it rests on, so a person can check it without replaying the recording.
03
Every recording, not a sample
A reviewer can listen to a few recordings a week. A workflow reads all of them, the same way each time.
04
An audit trail on every run
What went in, what came out, and which version of the function produced it. Open any result and trace it to its recording.
05
The same workflow as documents
A recording is one more input. Classify, Enrich, Decide, and Send work on it exactly as they do on a PDF.
06
Priced per call
A recording costs the same per function call as any other file. There is no per-minute meter.

08Questions

Questions about audio intelligence.

What is audio intelligence?: Audio intelligence is turning recordings into structured data a system can act on: who spoke, what was said, what was agreed, and what should happen next. With bem you define the record you want as a JSON schema, send the recording, and get that record back with an audit trail on the run.
What kinds of audio can bem read?: Any recording a person could listen to and take notes from. The examples on this page are an earnings call, an insurance claim call, a medical consultation, an air traffic control tape, and calls to a bank's phone line. The same approach covers voicemail, dictation, hearings, public meetings, and radio.
Which audio formats and file sizes does bem accept?: MP3, WAV, and M4A. A workflow call accepts files up to 50 MB each, sent as a multipart upload or as base64 JSON. The audio track of an MP4 or MOV video is read the same way.
Does bem transcribe the recording?: Yes. Speech is transcribed before extraction, and the record is filled in from what was said. If you want the transcript itself, ask for it in your schema.
How is this different from a transcription service?: A transcription service returns the words. bem returns the record: the figures from an earnings call as numbers, a claim as a first notice of loss, a call scored against your checks. The words come with it, as the evidence for each field.
Can bem score calls against our own quality scorecard?: Yes. Each line of your scorecard is a field in the schema, with a description of what counts as meeting it. Ask for the agent's words as well and every line that is met comes back with its quote. One of the calls on this page is scored that way.
How does it handle a queue with many kinds of recording?: A Classify function sorts each recording by what it is, and sends it to the Extract function for that kind. The queue on this page is 15 calls of 7 kinds, filed by one workflow.
How long does it take to process a recording?: In the recorded runs on this page, a single Extract function took between 3.7 and 12.7 seconds on recordings of 27 seconds to two and a half minutes. Each example shows the time its run took.
How is audio priced?: Per function call, the same as any other file. There is no per-minute meter, and the first 100 function calls are free.

See pricing Supported file types

Send your first recording.

The first 100 function calls are free. Pricing is per task, not per token or page. Pay as you go, with no lock-in.

Start building free Talk to an engineer See pricing

Audio intelligence for critical industries.

Input

Components

Decision

Investor days and conference talks

Central bank and regulator briefings

Recorded statements

Dictation and handover

Hearings and depositions

Public meetings

Radio and voice nets

Check calls and dispatch

Disclosures and promises to pay

Third-party verification

Voice notes from the job

Voicemail

Your schema, not ours

The words behind each field

Every recording, not a sample

An audit trail on every run

The same workflow as documents

Priced per call

What is audio intelligence?

What kinds of audio can bem read?

Which audio formats and file sizes does bem accept?

Does bem transcribe the recording?

How is this different from a transcription service?

Can bem score calls against our own quality scorecard?

How does it handle a queue with many kinds of recording?

How long does it take to process a recording?

How is audio priced?

Send your first recording.