whisper-asr v0.1.0

Whisper ASR Webservice — a self-hosted speech-to-text API powered by OpenAI's Whisper. Transcribe or translate audio to text over a simple HTTP endpoint, entirely on your own hardware. Deployed as a host-networked Nomad service with a persistent model-cache volume.

nomad-pack run whisper-asr --registry nomploy

…or one line (add registry + run)

curl -fsSL https://packs.nomploy.com/install.sh | sh -s -- whisper-asr

Needs nomad-pack on PATH. The script only adds the nomploy registry and runs this pack.

Source ↗ Project ↗ ★ 3.3k ⚑ Report an issue

1 task http 9000 1 volume image onerahmet/openai-whisper-asr-webservice:latest tracks :latest image bumped today

Variables 10

values.hcl

Save as values.hcl, edit, then run:

nomad-pack run whisper-asr -f values.hcl --registry nomploy

# The name of the Nomad job.
job_name = "whisper-asr"

# The Nomad namespace to deploy into.
namespace = "default"

# The datacenters to deploy to.
datacenters = ["*"]

# The Whisper ASR Webservice image. Pin a tag in production.
image = "onerahmet/openai-whisper-asr-webservice:latest"

# Host port for the ASR HTTP API. The container listens on 9000.
port = 9000

# Whisper model size (ASR_MODEL): tiny, base, small, medium, large-v3. Bigger = more accurate but slower and heavier.
asr_model = "base"

# Inference engine (ASR_ENGINE): openai_whisper, faster_whisper, or whisperx.
asr_engine = "faster_whisper"

# Named volume for the downloaded model cache (/root/.cache). Avoids re-downloading on restart.
cache_volume = "whisper_cache"

# Placement constraints. Pin to the node holding the volume. On a nomploy cluster: attribute = "$${meta.nomploy_control_plane}", operator = "=", value = "true".
constraints = []

# Resources for the Whisper task. Speech-to-text is CPU/RAM heavy; larger models need much more.
resources = {
    cpu    = 2000
    memory = 2048
  }

Name	Type	Default	Description
job_name	string	"whisper-asr"	The name of the Nomad job.
namespace	string	"default"	The Nomad namespace to deploy into.
datacenters	list	["*"]	The datacenters to deploy to.
image	string	"onerahmet/openai-whisper-asr-webservice:latest"	The Whisper ASR Webservice image. Pin a tag in production.
port	number	9000	Host port for the ASR HTTP API. The container listens on 9000.
asr_model	string	"base"	Whisper model size (ASR_MODEL): tiny, base, small, medium, large-v3. Bigger = more accurate but slower and heavier.
asr_engine	string	"faster_whisper"	Inference engine (ASR_ENGINE): openai_whisper, faster_whisper, or whisperx.
cache_volume	string	"whisper_cache"	Named volume for the downloaded model cache (/root/.cache). Avoids re-downloading on restart.
constraints	list	[]	Placement constraints. Pin to the node holding the volume. On a nomploy cluster: attribute = "$${meta.nomploy_control_plane}", operator = "=", value = "true".
resources	object	{ cpu = 2000 memory = 2048 }	Resources for the Whisper task. Speech-to-text is CPU/RAM heavy; larger models need much more.

Back up this pack

This pack stores data in one Docker named volume: whisper_cache

restic

# Run on the node hosting this pack. Point restic at your repo first:
#   export RESTIC_REPOSITORY="s3:https://<account>.r2.cloudflarestorage.com/<bucket>"
#   export RESTIC_PASSWORD="<repo-password>"
#   export AWS_ACCESS_KEY_ID=<key>  AWS_SECRET_ACCESS_KEY=<secret>
restic backup \
  /var/lib/docker/volumes/whisper_cache/_data

rclone (sync to S3/R2)

rclone sync /var/lib/docker/volumes/whisper_cache/_data backup:<bucket>/whisper_cache

Paths assume the default Docker volume location (/var/lib/docker/volumes). Restore by stopping the job, restoring files into the same volume, and re-running the pack.

Readme

whisper-asr

Whisper ASR Webservice — a self-hosted speech-to-text API powered by OpenAI's Whisper. Transcribe or translate audio to text over a simple HTTP endpoint, running entirely on your own hardware (great for subtitles, voice notes, and pipelines that shouldn't leave your network).

Single host-networked Nomad service with a persistent model-cache volume.

Deploy

nomad-pack registry add nomploy https://github.com/Nomploy/nomad-packs
nomad-pack run whisper-asr --registry=nomploy

Configure

Variable	Default	Description
`port`	`9000`	HTTP API port. The container listens on 9000.
`asr_model`	`base`	Whisper model (`ASR_MODEL`): `tiny`/`base`/`small`/`medium`/`large-v3`. Bigger = more accurate, slower, heavier.
`asr_engine`	`faster_whisper`	Engine (`ASR_ENGINE`): `openai_whisper`, `faster_whisper`, `whisperx`.
`cache_volume`	`whisper_cache`	`/root/.cache` — downloaded model cache.
`image`	`onerahmet/openai-whisper-asr-webservice:latest`	Container image. Pin a tag in production.
`resources`	`{ cpu = 2000, memory = 2048 }`	Task resources. Larger models need much more.

Transcribe with curl -F "audio_file=@sample.mp3" "http://<node-ip>:9000/asr?output=txt"; interactive docs at /docs. First boot downloads the model (needs internet). For real-time speed on big models use a GPU image/runtime. The API is unauthenticated — keep it internal or front it with a proxy. Pin the job to the node holding the volume with constraints.