Skip to navigation

AWS Marketplace (SageMaker)

Run Finch in your own AWS account as an Amazon SageMaker endpoint.

Murf Finch is available on AWS Marketplace as an Amazon SageMaker model package. You subscribe with your AWS account, deploy an endpoint in that account, and call it directly. There is no Murf API key, and no Murf server is in the request path. The container runs with SageMaker network isolation, so audio and transcripts never leave your endpoint. AWS bills you and pays Murf.

The listing has two versions. Finch Global transcribes English and six European languages. Finch Indic transcribes 11 Indian languages plus English (India), and returns word timestamps. Pick the version when you copy the model package ARN. The ARN is different for every version and every Region.

This listing is speech to text only. Text to speech is a separate product: Falcon 2.1 on AWS Marketplace. The hosted Speech to Text API is separate from this page: stream live audio over a WebSocket, or upload a file. Both use a Murf API key. This endpoint uses the SageMaker runtime, and its fields are the ones on this page. A SageMaker batch job is not the hosted file API.

Versions

VersionLanguagesWord timestampsTranscript text
Finch GlobalEnglish, German, Spanish, French, Italian, Portuguese, DutchEmpty words arrayLowercase, no punctuation
Finch IndicHindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Assamese, English (India)Start and end time for each wordPunctuated text

Each version transcribes only its own languages. Omit the language option and Finch detects the language from the audio. A language code from the other version returns HTTP 400.

Ways to call the endpoint

ModeAPIUse it for
FileInvokeEndpointOne audio file, up to 25 MB
Live streamInvokeEndpointWithBidirectionalStreamAudio as it is captured. Partials while the speaker talks, one final per turn
Batch transformCreateTransformJobA folder of audio files in S3, up to one hour each

Prerequisites

  • Subscribe. AWSMarketplaceManageSubscriptions, or the aws-marketplace:ViewSubscriptions, Subscribe, and Unsubscribe actions. Nothing is charged until an endpoint or a batch job runs.

  • Deploy. AmazonSageMakerFullAccess on a SageMaker execution role. SageMaker Studio and notebook instances already have a role you can use. To create one:

    aws iam create-role --role-name murf-finch-sagemaker \
    --assume-role-policy-document '{"Version":"2012-10-17","Statement":[{"Effect":"Allow","Principal":{"Service":"sagemaker.amazonaws.com"},"Action":"sts:AssumeRole"}]}'
    aws iam attach-role-policy --role-name murf-finch-sagemaker \
    --policy-arn arn:aws:iam::aws:policy/AmazonSageMakerFullAccess
  • GPU quota. In Service Quotas → Amazon SageMaker, set ml.g5.2xlarge for endpoint usage to at least 1. For batch transform, ml.g6.2xlarge for transform job usage is often 0 until you request it. Quota increases can take a day or more.

    aws service-quotas list-service-quotas --service-code sagemaker --region us-east-1 \
    --query "Quotas[?contains(QuotaName,'ml.g5.2xlarge for endpoint usage')].[QuotaName,Value]"
  • Live streaming from Python. Python 3.12+ and aws-sdk-sagemaker-runtime-http2[awscrt]>=0.11. boto3 does not support the bidirectional API.

Quickstart

About 15 minutes, most of it waiting for the endpoint to start. The notebook does the same steps. It reads an ARN for the notebook’s Region from model_packages.json when that Region is listed there. For any other Region, or to pin a version, paste the ARN from the listing into MODEL_PACKAGE_ARN.

1

Subscribe and copy the ARN

Open the listing, choose Continue to subscribe, review pricing and the EULA, then Accept offer. Choose Continue to configuration, pick Finch Global or Finch Indic and your Region, and copy the product ARN.

Already subscribed? AWS Marketplace → Manage subscriptions → Murf Finch → Launch new configuration, select the version and Region, and copy the ARN.

2

Deploy the endpoint

Network isolation is required for Marketplace model packages. The startup health check needs 900 seconds because the model takes a few minutes to load. Least-outstanding-requests routing sends each new stream to the least busy instance once you run more than one.

import boto3, time
region = "us-east-1" # the Region you copied the ARN for
model_package_arn = "<model package ARN from the listing>"
role = "arn:aws:iam::<account-id>:role/murf-finch-sagemaker"
name = f"murf-finch-{time.strftime('%Y%m%d-%H%M%S')}"
sm = boto3.client("sagemaker", region_name=region)
sm.create_model(
ModelName=name,
ExecutionRoleArn=role,
PrimaryContainer={"ModelPackageName": model_package_arn},
EnableNetworkIsolation=True,
)
sm.create_endpoint_config(
EndpointConfigName=name,
ProductionVariants=[{
"VariantName": "AllTraffic",
"ModelName": name,
"InstanceType": "ml.g5.2xlarge",
"InitialInstanceCount": 1,
"ContainerStartupHealthCheckTimeoutInSeconds": 900,
"RoutingConfig": {"RoutingStrategy": "LEAST_OUTSTANDING_REQUESTS"},
}],
)
sm.create_endpoint(EndpointName=name, EndpointConfigName=name)
sm.get_waiter("endpoint_in_service").wait(
EndpointName=name, WaiterConfig={"Delay": 30, "MaxAttempts": 80}
)

The endpoint is ready in about 8–10 minutes. describe_model_package and create_model fail when the account is not subscribed to that version.

3

Transcribe a file

The body is the audio file. samples/en_us_sample.wav is an English clip for Finch Global, and samples/hi_in_sample.wav is Hindi for Finch Indic. Both are from the FLEURS dataset (CC BY 4.0).

import json
runtime = boto3.client("sagemaker-runtime", region_name=region)
resp = runtime.invoke_endpoint(
EndpointName=name,
ContentType="audio/wav",
Body=open("en_us_sample.wav", "rb").read(),
)
print(json.loads(resp["Body"].read())["text"])

Or from a terminal, with the endpoint name in NAME:

aws sagemaker-runtime invoke-endpoint --endpoint-name "$NAME" \
--content-type audio/wav --body fileb://samples/en_us_sample.wav out.json && cat out.json
4

Delete the endpoint when you are done

You pay AWS for the GPU instance while the endpoint exists, whether or not you send requests. Delete test endpoints when you finish.

sm.delete_endpoint(EndpointName=name)
sm.delete_endpoint_config(EndpointConfigName=name)
sm.delete_model(ModelName=name)

The subscription stays active, so you can deploy again later. To end it, delete your models first, then cancel under AWS Marketplace → Manage subscriptions.

Transcribe a file

InvokeEndpoint returns the full transcript. WAV, FLAC, MP3, OGG/Opus, M4A, and other common formats are accepted. A request can be up to 25 MB (about 13 minutes of 16 kHz WAV, and longer for MP3 or Opus) and must finish within SageMaker’s 60-second limit. Longer files use batch transform.

Content-Type is audio/wav, audio/mpeg, application/octet-stream, or the type that matches the file.

Options go in the X-Amzn-SageMaker-Custom-Attributes header (CustomAttributes in boto3, --custom-attributes in the CLI) as comma-separated key=value pairs. They are not fields in a JSON body.

OptionRequiredDescription
language_codeNoLanguage of the audio, such as en or de on Finch Global and hi on Finch Indic. Detected when omitted.
timestamps_granularityNoword (default) or none. Finch Indic returns word start and end times. Finch Global returns an empty words array either way.
keytermsNoWords or phrases to recognize more reliably. Separate phrases with a vertical bar.
resp = runtime.invoke_endpoint(
EndpointName=name,
ContentType="audio/wav",
CustomAttributes="language_code=hi,timestamps_granularity=word,keyterms=Murf|Finch",
Body=open("hi_in_sample.wav", "rb").read(),
)

Response

Every response includes these fields:

FieldDescription
textThe transcript. Finch Global is lowercase and has no punctuation.
language_codeDetected or requested language.
language_probabilityConfidence of the language detection, from 0 to 1.
wordsFinch Indic: each word as text, start, end (seconds), and type. Empty for Finch Global, and when timestamps_granularity is none.
audio_duration_secsLength of the audio. This is the billed duration.
transcription_idUnique ID. Include it in support requests.
{
"transcription_id": "88167718-7e7a-4529-992c-2aee87b56a04",
"language_code": "en",
"language_probability": 1.0,
"text": "he did not set a figure for the cuts saying they would be made based on china's economic output",
"words": [],
"audio_duration_secs": 6.36
}

A Finch Indic word looks like this:

{"text": "यह", "start": 0.788, "end": 0.958, "type": "word"}

Stream live audio

Open a bidirectional stream with InvokeEndpointWithBidirectionalStream. examples/stream.py is a complete Python client. It uses AWSCRTHTTPClient against the runtime endpoint on port 8443. Another transport raises UnsupportedTransportError.

pip install "aws-sdk-sagemaker-runtime-http2[awscrt]>=0.11" boto3
python examples/stream.py --endpoint "$NAME" --region "$AWS_REGION" samples/en_us_sample.wav

--language de sets the language, --bvc suppresses background voices, and --raw prints every message.

Options go in ModelQueryString, for example language=hi&keywords=Murf,Finch. Names here differ from a file request: the query uses language and keywords. The file header uses language_code and keyterms.

OptionDescription
languageOne code (en, hi, de), or a comma-separated set to detect from. Detected when omitted.
keywordsComma-separated words or phrases to recognize more reliably.
commit_modevad (default) ends a turn after a short silence. manual ends a turn only when you send Finalize.
vad_silence_threshold_secsSilence that ends a turn in vad mode. Default 0.55, range 0.192 to 3.0. Raise it for slow speakers, lower it for snappier agents.
bvctrue suppresses background voices, such as a TV or people nearby. Off by default. The capacity figures below were measured with it off.

You send:

  1. Optionally, one JSON text message describing the audio. Without it the endpoint expects 16-bit little-endian mono PCM at 16 kHz.

    {"type": "AudioConfig", "codec": "pcm_s16le", "sample_rate": 16000, "channels": 1}

    codec is pcm_s16le, mulaw, alaw, or opus (one packet per frame). sample_rate is 8000, 12000, 16000, 24000, 32000, 44100, or 48000.

  2. Audio as binary frames, ideally 20–100 ms each, at real-time pace.

  3. Control messages as JSON text. Finalize ends the current turn and keeps the stream open. CloseStream flushes and closes. KeepAlive holds a paused stream open; send it instead of silence.

    {"type": "Finalize"}

You receive JSON text messages:

{"type": "SessionStarted", "session_id": "20261112T103817Z_05a4...", "language": null}
{"type": "partial", "transcript": "he did not set a figure", "rawEvent": {"event": "Update", "turn_index": 0}}
{
"type": "final",
"transcript": "he did not set a figure for the cuts saying they will be made based on china's economic output",
"rawEvent": {
"event": "EndOfTurn",
"turn_index": 0,
"audio_window_start": 0.916,
"audio_window_end": 5.26,
"language": "en",
"language_confidence": 1.0,
"words": []
}
}
  • partial updates the current turn as the speaker talks. Show it as live captions.
  • One final per turn carries the settled text. A final with an empty transcript means the turn contained no speech.
  • rawEvent adds timing in seconds from the start of the stream, the detected language, and, on Finch Indic, word timestamps.
  • The transcript field on a stream is transcript. A file response uses text.
  • Errors arrive as a message with "type": "Error" and a message string.

A stream closes after 10 seconds without a first message, or after 180 seconds without audio or KeepAlive.

Batch transform

Put one audio file per S3 object under an input prefix. Each output object is the input name plus .out and holds the same JSON as a real-time response. Files can be up to one hour long.

aws sagemaker create-transform-job --transform-job-name finch-batch-1 --model-name "$NAME" \
--transform-input '{"DataSource":{"S3DataSource":{"S3DataType":"S3Prefix","S3Uri":"s3://my-bucket/audio/"}},"ContentType":"audio/wav","SplitType":"None"}' \
--transform-output '{"S3OutputPath":"s3://my-bucket/transcripts/","Accept":"application/json"}' \
--transform-resources '{"InstanceType":"ml.g6.2xlarge","InstanceCount":1}' \
--max-payload-in-mb 100 --batch-strategy SingleRecord

SplitType None keeps each object as one audio file. Batch transform runs on ml.g6.2xlarge, ml.g6.xlarge, and ml.g5.xlarge, and needs “for transform job usage” quota for that type. The notebook uploads the sample clips and prints each transcript.

Instances and performance

Measured on Amazon SageMaker in us-east-2 in November 2026. Live 16 kHz streams of FLEURS speech were sent at real-time pace, 120 seconds per load level. “Streams” is the highest load tested at which 90% of final transcripts arrived within 250 ms of the end of speech. The stream counts are for Finch Global. Finch Indic stayed under 165 ms (p90) at 64 streams on every instance below. A single stream is 45–85 ms (p90) on all of them.

InstanceGPULive streamsFinal transcript, p90Best for
ml.g5.2xlarge (recommended)NVIDIA A10G96222 msLowest cost per stream
ml.g6.xlargeNVIDIA L464217 msLowest price per instance
ml.g6.2xlargeNVIDIA L464204 ms
ml.g5.xlargeNVIDIA A10G64143 ms
ml.g6e.xlargeNVIDIA L40S9690 msLowest latency

On one ml.g5.2xlarge, p90 final latency is 66 ms at 1 stream, 106 ms at 32, 150 ms at 64, and 222 ms at 96.

Size the endpoint as peak concurrent streams divided by streams per instance, rounded up, plus one instance for redundancy. For variable traffic, add SageMaker auto scaling on ConcurrentRequestsPerModel.

Other instance types are not supported in this version. ml.g4dn is too slow for live use, and ml.g7e does not start.

When a GPU type is unavailable in an Availability Zone, give the endpoint a fallback list with instance pools. Replace InstanceType in the production variant with:

"InstancePools": [
{"InstanceType": "ml.g5.2xlarge", "Priority": 1},
{"InstanceType": "ml.g6.xlarge", "Priority": 2},
{"InstanceType": "ml.g6.2xlarge", "Priority": 3}
],
"VariantInstanceProvisionTimeoutInSeconds": 300

Every type in the pool needs quota of at least 1.

Accuracy

Word error rate on the FLEURS test sets, with text normalized before scoring:

LanguageWERLanguageWER
English (US)5.2%Italian3.4%
German5.5%Dutch6.5%
Spanish (Latin America)2.6%Portuguese (Brazil)5.6%
French5.5%Finch Global average4.9%
Hindi (Finch Indic)10.3%

Accuracy is lower on noisy, overlapping, or far-field speech. Test with your own audio.

Limitations

  • Speaker diarization is not included.
  • Finch Global returns lowercase text without punctuation, and its words array is empty.
  • Each version transcribes only the languages in the versions table.

Pricing

Real-time and streaming usage is billed per minute of audio, in fractions of a minute. Streams of PCM, mu-law, or A-law are billed by the audio duration sent. Opus streams are billed by session time. Pauses covered by KeepAlive are not billed for PCM streams. For a file request, audio_duration_secs is the billed duration.

Batch transform is billed per instance-hour while the job runs. AWS separately bills the SageMaker instance hours. Current prices are on the listing.

Troubleshooting

SymptomFix
ResourceLimitExceeded when creating the endpoint or transform jobRequest a quota increase for that instance type, or use another type from the table above.
Endpoint Failed with InsufficientInstanceCapacityAWS has no free GPUs of that type in the Region right now. Retry, use an instance pool, or another Region.
Endpoint fails its health checkKeep ContainerStartupHealthCheckTimeoutInSeconds at 900, and use a supported instance type. ml.g4dn and ml.g7e are not supported.
ValidationError mentioning network isolationCreate the model with EnableNetworkIsolation set to true.
HTTP 413 on InvokeEndpointThe file is over 25 MB. Compress it, split it, or use batch transform.
ModelError with status 400The audio could not be decoded, or language_code is not in this version. The message says which.
Stream fails with UnsupportedTransportErrorUse smithy_http.aio.crt.AWSCRTHTTPClient as the transport, as in examples/stream.py.
Stream returns nothingThe audio format does not match AudioConfig. The default is 16 kHz 16-bit mono PCM. Silence produces an empty final transcript.
Transcript in the wrong languageSet language_code on a file, or language on a stream. Finch Global does not transcribe Indian languages. Deploy Finch Indic for those.

Security and privacy

  • Audio and transcripts stay in your account. With network isolation, the container has no network access and sends nothing to Murf or any third party.
  • Audio and transcripts are not used to train models.
  • The endpoint writes one line per request to CloudWatch Logs with status and timing. Audio and transcript text are never logged.

Support

Murf cannot see your requests, audio, or transcripts. Email support@murf.ai with the AWS Region, instance type, listing version, the transcription_id or session_id, and the error message. Billing disputes go through AWS Marketplace.