welsh-whisperx

Gweinydd lleferydd-i-destun Cymraeg wedi'i adeiladu ar WhisperX ac wedi'i optimeiddio ar gyfer cywirdeb a chefnogaeth i'r iaith Gymraeg.

A production-ready Welsh speech-to-text server built on WhisperX, optimized for accuracy, scalability, and Welsh language support.

English version below / Fersiwn Saesneg isod

Prif Fanteision

🎯 Wedi'i Optimeiddio ar gyfer y Gymraeg

Modelau wedi'u mireinio: Modelau Whisper a Wav2Vec2 penodol i'r Gymraeg gan Techiaith
Normaleiddio testun Cymraeg: Normaleiddio testun llafar yn awtomatig
- Ehangu talfyriadau (ma'→mae, ca'l→cael)
- Safoni amrywiadau tafodieithol (petha→pethau, gwbod→gwybod)
- Normaleiddio rhagenwau a ffurfiau berfol
Canfod iaith: Canfod Cymraeg/Saesneg yn awtomatig gyda Lingua
Aliniad cywir: Stampiau amser ar lefel geiriau wedi'u hoptimeiddio ar gyfer seineg y Gymraeg

⚡ Perfformiad Uchel a Graddadwyedd

Cyflymu GPU: Cefnogaeth lawn NVIDIA CUDA ar gyfer trawsgrifio cyflym
Gweithwyr ciw pwrpasol: Gweithwyr blaenoriaeth uchel a rhagosodedig ar wahân, yn raddadwy'n annibynnol
Blaenoriaethu tasgau: System flaenoriaethu dwy lefel (llwybro ciwiau + blaenoriaethau rhifol)
Swpio effeithlon: Casgliad WhisperX ar gyfer cyflymu 12x
Rheoli cof: Glanhau awtomatig ac ailgylchu gweithwyr i atal gollyngiadau cof

🔧 Yn Barod ar gyfer Cynhyrchu

Defnyddio mewn cynwysyddion: Docker Compose ar gyfer amgylcheddau GPU a CPU
Monitro iechyd: Gwiriadau iechyd cynhwysfawr ar gyfer yr holl wasanaethau
Glanhau awtomatig: Dileu ffeiliau hen yn ôl amserlen i atal gormod o ddata ar y ddisg
Parhauster tasgau: Ciw wedi'i gefnogi gan Redis yn goroesi ailgychwyn
Logio: Logio fesul tasg ar gyfer dadfygio a monitro
Dogfennaeth API: Dogfennau OpenAPI/Swagger wedi'u cynhyrchu'n awtomatig

📊 Fformatau Allbwn Cyfoethog

JSON: Trawsgrifiad llawn gyda stampiau amser ar lefel geiriau a sgoriau
SRT/VTT: Ffeiliau isdeitlau gyda rhannu llinellau deallus
ELAN: Anodiadau llinell amser ar gyfer dadansoddi ieithyddol
Testun: Trawsgrifiad testun plaen
Diarization siaradwyr: Adnabod siaradwyr gwahanol (SIARADWR_01, ayyb.)
Testun wedi'i normaleiddio: Normaleiddio testun Cymraeg wedi'i gynnwys

🚀 Cyfeillgar i Ddatblygwyr

API RESTful: Diweddbwyntiau HTTP syml ar gyfer yr holl weithrediadau
Archwilio ciwiau: Monitro tasgau aros a statws gweithwyr
Defnyddio hyblyg: Modd CPU yn unig ar gyfer datblygu, GPU ar gyfer cynhyrchu
Rheoli amgylchedd: Ffurfweddiad tryloyw gyda gwahanu cyfrinachau
Dogfennaeth gynhwysfawr: Canllawiau pensaernïaeth, defnyddio, a datrys problemau

🔒 Nodweddion Menter

Cefnogaeth modelau preifat: Dilysu gyda HuggingFace ar gyfer modelau preifat
Tracio tasgau: IDau tasg parhaol ar draws ailgychwyn y gweinydd
Cadw canlyniadau: Terfyn 7 diwrnod ar ganlyniadau gyda Redis
CORS wedi'i alluogi: Mynediad API cyfeillgar i borwyr
Trin gwallau: Dirywiad gosgeiddig a negeseuon gwall manwl

A production-ready Welsh speech-to-text server built on WhisperX, optimized for accuracy, scalability, and Welsh language support.

English version below / Fersiwn Saesneg isod

Prif Fanteision

🎯 Wedi'i Optimeiddio ar gyfer y Gymraeg

Modelau wedi'u mireinio: Modelau Whisper a Wav2Vec2 penodol i'r Gymraeg gan Techiaith
Normaleiddio testun Cymraeg: Normaleiddio testun llafar yn awtomatig
- Ehangu talfyriadau (ma'→mae, ca'l→cael)
- Safoni amrywiadau tafodieithol (petha→pethau, gwbod→gwybod)
- Normaleiddio rhagenwau a ffurfiau berfol
Canfod iaith: Canfod Cymraeg/Saesneg yn awtomatig gyda Lingua
Aliniad cywir: Stampiau amser ar lefel geiriau wedi'u hoptimeiddio ar gyfer seineg y Gymraeg

⚡ Perfformiad Uchel a Graddadwyedd

Cyflymu GPU: Cefnogaeth lawn NVIDIA CUDA ar gyfer trawsgrifio cyflym
Gweithwyr ciw pwrpasol: Gweithwyr blaenoriaeth uchel a rhagosodedig ar wahân, yn raddadwy'n annibynnol
Blaenoriaethu tasgau: System flaenoriaethu dwy lefel (llwybro ciwiau + blaenoriaethau rhifol)
Swpio effeithlon: Casgliad WhisperX ar gyfer cyflymu 12x
Rheoli cof: Glanhau awtomatig ac ailgylchu gweithwyr i atal gollyngiadau cof

🔧 Yn Barod ar gyfer Cynhyrchu

Defnyddio mewn cynwysyddion: Docker Compose ar gyfer amgylcheddau GPU a CPU
Monitro iechyd: Gwiriadau iechyd cynhwysfawr ar gyfer yr holl wasanaethau
Glanhau awtomatig: Dileu ffeiliau hen yn ôl amserlen i atal gormod o ddata ar y ddisg
Parhauster tasgau: Ciw wedi'i gefnogi gan Redis yn goroesi ailgychwyn
Logio: Logio fesul tasg ar gyfer dadfygio a monitro
Dogfennaeth API: Dogfennau OpenAPI/Swagger wedi'u cynhyrchu'n awtomatig

📊 Fformatau Allbwn Cyfoethog

JSON: Trawsgrifiad llawn gyda stampiau amser ar lefel geiriau a sgoriau
SRT/VTT: Ffeiliau isdeitlau gyda rhannu llinellau deallus
ELAN: Anodiadau llinell amser ar gyfer dadansoddi ieithyddol
Testun: Trawsgrifiad testun plaen
Diarization siaradwyr: Adnabod siaradwyr gwahanol (SIARADWR_01, ayyb.)
Testun wedi'i normaleiddio: Normaleiddio testun Cymraeg wedi'i gynnwys

🚀 Cyfeillgar i Ddatblygwyr

API RESTful: Diweddbwyntiau HTTP syml ar gyfer yr holl weithrediadau
Archwilio ciwiau: Monitro tasgau aros a statws gweithwyr
Defnyddio hyblyg: Modd CPU yn unig ar gyfer datblygu, GPU ar gyfer cynhyrchu
Rheoli amgylchedd: Ffurfweddiad tryloyw gyda gwahanu cyfrinachau
Dogfennaeth gynhwysfawr: Canllawiau pensaernïaeth, defnyddio, a datrys problemau

🔒 Nodweddion Menter

Cefnogaeth modelau preifat: Dilysu gyda HuggingFace ar gyfer modelau preifat
Tracio tasgau: IDau tasg parhaol ar draws ailgychwyn y gweinydd
Cadw canlyniadau: Terfyn 7 diwrnod ar ganlyniadau gyda Redis
CORS wedi'i alluogi: Mynediad API cyfeillgar i borwyr
Trin gwallau: Dirywiad gosgeiddig a negeseuon gwall manwl

Rhagofynion

Docker a Docker Compose (gofynnol)
Linux neu macOS — heb ei brofi ar Windows
Modd GPU: GPU NVIDIA + nvidia-docker2
Cof: Isafswm 8GB RAM (16GB+ yn argymhellir ar gyfer modd GPU)
Disg: O leiaf 20GB o le rhydd ar gyfer modelau a recordiadau

Cychwyn Cyflym

Gosod Cychwynnol

# Clonio'r ystorfa
git clone https://storfa.techiaith.cymru/lleferydd/stt-cy/whisperx-server
cd whisperx-server

# Gosod ffeiliau amgylchedd
make setup

# Golygu .env.secrets ac ychwanegu eich tocyn HuggingFace
# Cewch eich tocyn o: https://huggingface.co/settings/tokens
nano .env.secrets

# Ail-gynhyrchu .env gyda'ch tocyn
make setup

Defnyddio GPU (Argymhellir ar gyfer Cynhyrchu)

# Adeiladu a rhedeg gyda chefnogaeth GPU (angen nvidia-docker)
# Yn cychwyn 1 gweithiwr blaenoriaeth uchel + 1 gweithiwr rhagosodedig
make build-gpu

# Graddio gweithwyr yn annibynnol
make scale-high-gpu N=2       # 2 weithiwr blaenoriaeth uchel
make scale-default-gpu N=3    # 3 gweithiwr rhagosodedig

Defnyddio CPU (Ar gyfer Datblygu/Profi)

# Adeiladu a rhedeg gyda CPU yn unig (dim GPU yn ofynnol)
# Yn cychwyn 1 gweithiwr blaenoriaeth uchel + 1 gweithiwr rhagosodedig
make build-cpu

# Graddio gweithwyr yn annibynnol
make scale-high-cpu N=2       # 2 weithiwr blaenoriaeth uchel
make scale-default-cpu N=2    # 2 weithiwr rhagosodedig

Gorchmynion Cyffredin

make help          # Dangos yr holl orchmynion sydd ar gael
make ready         # Aros i'r modelau orffen llwytho
make status        # Dangos cynwysyddion sy'n rhedeg
make health        # Gwirio iechyd gwasanaeth
make logs          # Gweld logiau cynwysyddion
make down          # Stopio'r holl gynwysyddion

Mynediad i'r API

Unwaith y bydd yn rhedeg, mynediad i'r API yn: http://localhost:5511

Gwiriad Iechyd: http://localhost:5511/health/ Dogfennau API: http://localhost:5511/docs (wedi'i gynhyrchu'n awtomatig gan FastAPI)

Ap Gwe

Mae ap gwe syml ar gael yn y ffolder app/ ar gyfer cyflwyno trawsgrifiadau a chyfieithiadau trwy'r porwr. I'w ddefnyddio:

python3 -m http.server 8080 --directory app/

Yna agorwch http://localhost:8080 yn eich porwr. Mae'r ap yn cysylltu â'r API ar localhost:5511.

Enghreifftiau Defnyddio'r API

Trawsgrifio Sain (Ffurf Fer, Cydamserol)

# Rhagosodiad: blaenoriaeth uchel (wedi'i lwybro i giw high_priority)
curl -X POST "http://localhost:5511/transcribe/" \
  -F "soundfile=@sain.wav"

# Blaenoriaeth normal yn benodol (dal yn mynd i giw high_priority)
curl -X POST "http://localhost:5511/transcribe/" \
  -F "soundfile=@sain.wav" \
  -F "priority=normal"

Trawsgrifio Sain Hir (Anghydamserol)

# Cyflwyno tasg (wedi'i lwybro i giw default)
curl -X POST "http://localhost:5511/transcribe_long_form/" \
  -F "soundfile=@sain_hir.wav"

# Dychwelyd: {"id": "uuid", "version": 2, "success": true}

# Gwirio statws
curl "http://localhost:5511/get_status/?stt_id=<uuid>"

# Cael canlyniadau pan fo'n gyflawn
curl "http://localhost:5511/get_json/?stt_id=<uuid>"
curl "http://localhost:5511/get_srt/?stt_id=<uuid>"
curl "http://localhost:5511/get_vtt/?stt_id=<uuid>"

Cyfieithu Sain (Cymraeg → Saesneg)

# Ffurf fer (cydamserol, uchafswm 480KB)
curl -X POST "http://localhost:5511/translate/" \
  -F "soundfile=@sain.wav"

# Ffurf hir (anghydamserol, unrhyw faint)
curl -X POST "http://localhost:5511/translate_long_form/" \
  -F "soundfile=@sain_hir.wav"

Alinio Testun â Sain (Stampiau Amser fesul Gair)

# Ffurf fer (cydamserol, uchafswm 480KB)
curl -X POST "http://localhost:5511/align/" \
  -F "soundfile=@sain.wav" \
  -F "text=Mae ganddynt dau o blant, mab a merch"

# Ffurf hir (anghydamserol, unrhyw faint — yn rhedeg ar CPU)
curl -X POST "http://localhost:5511/align_long_form/" \
  -F "soundfile=@sain_hir.wav" \
  -F "text=Testun llawn i'w alinio â'r sain..."

# Gwirio statws
curl "http://localhost:5511/get_status/?stt_id=<uuid>"

# Cael canlyniadau
curl "http://localhost:5511/get_json/?stt_id=<uuid>"

Mewnbwn Bysellfwrdd Amser Real (Blaenoriaeth Uchel Bob Amser)

curl -X POST "http://localhost:5511/keyboard/" \
  -F "audio_file=@mewnbwn_llais.wav"
# Wedi'i lwybro bob amser i giw high_priority gyda blaenoriaeth frys

Pensaernïaeth

Cydrannau

┌──────────────────┐
│    FastAPI App    │  Porth 5511
│ (Gweinydd API)   │
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│  Brocer Redis    │  Porth 6379
│  (Ciw Tasgau)    │
├──────────────────┤
│ high_priority    │  /transcribe/, /keyboard/, /translate/
│ default          │  /transcribe_long_form/, /translate_long_form/
│ alignment        │  /align/, /align_long_form/
└──┬───┬───┬───┬───┘
   │   │   │   │
   ▼   ▼   ▼   ▼
┌────────┐ ┌──────────┐ ┌─────────┐ ┌───────────┐
│worker  │ │worker    │ │worker   │ │worker     │  Yn Raddadwy'n
│-high   │ │-high-2   │ │-default │ │-alignment │  Annibynnol
│GPU 1   │ │GPU 0     │ │GPU 1    │ │(CPU x2)   │
│(GPU 1) │ │(GPU 0)   │ │(GPU 1)  │ │           │
└────────┘ └──────────┘ └─────────┘ └───────────┘

Aseiniad GPU

Gwasanaeth	Ciw	GPU	Nodiadau
`worker-high`	`high_priority`	GPU 1	Yn rhannu GPU gyda worker-default
`worker-high-2`	`high_priority`	GPU 0	GPU pwrpasol, gorlif blaenoriaeth uchel
`worker-default`	`default`	GPU 1	Yn rhannu GPU gyda worker-high
`worker-alignment` (x2)	`alignment`	CPU	Alinio wav2vec2 ar CPU

Pan fo worker-default yn brysur gyda thrawsgrifiad hir, mae worker-high-2 ar GPU 0 yn sicrhau bod ceisiadau blaenoriaeth uchel yn dal i gael eu gwasanaethu'n gyflym.

Gwasanaethau

Application (FastAPI): Gweinydd API REST yn trin ceisiadau
worker-high (Celery): Pwrpasol ar gyfer tasgau blaenoriaeth uchel (/transcribe/, /keyboard/, /translate/) — GPU 1
worker-high-2 (Celery): Ail weithiwr blaenoriaeth uchel ar GPU pwrpasol (/transcribe/, /keyboard/, /translate/) — GPU 0
worker-default (Celery): Pwrpasol ar gyfer tasgau rhagosodedig (/transcribe_long_form/, /translate_long_form/) — GPU 1
worker-alignment (Celery): Pwrpasol ar gyfer alinio ar CPU (/align/, /align_long_form/) — 2 replica rhag ofn bod tasg hir yn rhwystro
Redis: Brocer negeseuon a storfa canlyniadau

System Ciwiau a Blaenoriaethu Tasgau

Mae'r system yn defnyddio blaenoriaethu dwy lefel: gweithwyr ciw pwrpasol + blaenoriaethau rhifol o fewn pob ciw.

Llwybro Ciwiau fesul Diweddbwynt

Diweddbwynt	Ciw	Blaenoriaeth Ddiofyn	Nodiadau
`/keyboard/`	`high_priority`	0 (brys)	Bob amser yn frys, dim paramedr blaenoriaeth
`/transcribe/`	`high_priority`	0 (uchel)	Yn derbyn `high` neu `normal`
`/translate/`	`high_priority`	0 (uchel)	Cyfieithu Cymraeg → Saesneg
`/transcribe_long_form/`	`default`	5 (normal)	Anghydamserol, dim paramedr blaenoriaeth
`/translate_long_form/`	`default`	5 (normal)	Cyfieithu anghydamserol, dim paramedr blaenoriaeth
`/align/`	`alignment`	0 (brys)	Alinio testun â sain, stampiau amser fesul gair (CPU)
`/align_long_form/`	`alignment`	5 (normal)	Alinio sain hir yn anghydamserol (CPU)

Aseiniad Ciwiau Gweithwyr

Mae gweithwyr yn cael eu neilltuo i giwiau penodol trwy'r newidyn amgylchedd WORKER_QUEUES:

Gweithiwr	Ciwiau	Dyfais	Yn Gwasanaethu
`worker-high`	`high_priority`	GPU 1	`/transcribe/`, `/keyboard/`, `/translate/`
`worker-high-2`	`high_priority`	GPU 0	`/transcribe/`, `/keyboard/`, `/translate/`
`worker-default`	`default`	GPU 1	`/transcribe_long_form/`, `/translate_long_form/`
`worker-alignment` (x2)	`alignment`	CPU	`/align/`, `/align_long_form/`

Blaenoriaethau Rhifol (o fewn pob ciw)

Gwerth	Enw	Defnyddir Gan
0	brys	`/keyboard/`, `/transcribe/` gyda `priority=high`, `/translate/`, `/align/`
5	normal	`/transcribe/` gyda `priority=normal`, `/transcribe_long_form/`, `/translate_long_form/`, `/align_long_form/`

Fformat Allbwn

Ymateb JSON

{
  "id": "id_sain",
  "version": 2,
  "success": true,
  "language": "cy",
  "segments": [
    {
      "audio_id": "hash_md5",
      "start": 0.0,
      "end": 2.5,
      "text": "Mae hyn yn brawf o drawsgrifio",
      "normalized": "Mae hyn yn brawf o drawsgrifio",
      "score": 0.95,
      "words": [
        {"word": "Mae", "start": 0.0, "end": 0.3, "score": 0.98},
        {"word": "hyn", "start": 0.3, "end": 0.5, "score": 0.96}
      ],
      "chars": [...]
    }
  ]
}

Nodweddion

Normaleiddio testun: Mae pob segment yn cynnwys text gwreiddiol a normalized (Cymraeg safonol)
Stampiau amser ar lefel geiriau: Amseru manwl gywir ar gyfer pob gair
Sgoriau hyder: Sgoriau cywirdeb fesul gair a fesul segment
Diarization siaradwyr: Labeli siaradwyr (SIARADWR_01, SIARADWR_02, ayyb.)
Canfod iaith: Adnabod Cymraeg/Saesneg yn awtomatig

Ffurfweddiad yr Amgylchedd

Mae'r prosiect hwn yn defnyddio dull hybrid ar gyfer newidynnau amgylchedd:

Strwythur Ffeiliau

config.env - Ffurfweddiad tryloyw (wedi'i gomitio i git)
- Enwau modelau, gosodiadau Redis, rhagosodiadau an-sensitif
- Golygwch y ffeil hon i newid ffurfweddiadau modelau
.env.secrets - Gwerthoedd sensitif (wedi'i gitignore)
- Tocynnau HuggingFace, cyfrineiriau
- Crëwch o .env.secrets.example
.env - Ffeil wedi'i chynhyrchu (wedi'i gitignore)
- Wedi'i chreu'n awtomatig gan make setup
- Yn cyfuno config.env + .env.secrets

Prif Opsiynau Ffurfweddu

# Ffurfweddiad Model (config.env)
WHISPER_MODEL_NAME=DewiBrynJones/whisper-large-v2-ft-btb-cv-cvad-ca-wlga-cy-ct2-2511
WHISPER_MODEL_LANGUAGE=cy
WAV2VEC2_MODEL=techiaith/wav2vec2-btb-cv-ft-cv-cy

# Glanhau Ffeiliau (config.env)
CLEANUP_ENABLED=true
FILE_RETENTION_DAYS=14
CLEANUP_SCHEDULE=0 2 * * *  # 2 AM yn ddyddiol

# Cyfrinachau (.env.secrets)
HF_AUTH_TOKEN=eich_tocyn_huggingface_yma

Newid Ffurfweddiad

I newid gosodiadau model:

# Golygu'r ffeil ffurfweddiad tryloyw
nano config.env

# Ail-gynhyrchu .env
make setup

I ddiweddaru'ch tocyn HuggingFace:

# Golygu ffeil gyfrinachau
nano .env.secrets

# Ail-gynhyrchu .env
make setup

Dogfennaeth

Dogfennaeth Graidd

CPU_GPU_DEPLOYMENT.md - Canllaw llawn ar gyfer defnyddio CPU vs GPU
CELERY_ARCHITECTURE.md - Deall y system ciw tasgau
DOCKER_HEALTHCHECKS.md - Monitro iechyd cynwysyddion

Dogfennaeth Nodweddion

FILE_CLEANUP.md - Glanhau ffeiliau trawsgrifio hen yn awtomatig
TASK_PRIORITIZATION.md - Llwybro ciwiau a system flaenoriaethu
IMPROVING_ALIGNMENT.md - Gwella cywirdeb stampiau amser ar lefel geiriau

Graddio a Defnyddio

GPU_SCALING.md - Terfynau cof GPU a ffurfweddiadau aml-GPU
ADVANCED_SCALING.md - Strategaethau graddio uwch
INFERENCE_DEPLOYMENT_ALTERNATIVES.md - Cymhariaeth o atebion defnyddio casgliad

Defnyddio Brodorol macOS

macos/README.md - Defnyddio Apple Silicon brodorol gyda chyflymu MLX
macos/HYBRID_MLX_ARCHITECTURE.md - Dogfennaeth pensaernïaeth hybrid MLX

Dogfennaeth Dechnegol

claude_refactors_2601.md - Gwelliannau pensaernïol diweddar

Graddio

Graddio Gweithwyr yn Annibynnol

Graddiwch weithwyr blaenoriaeth uchel a rhagosodedig yn annibynnol yn seiliedig ar eich llwyth gwaith:

# Gweithwyr GPU
make scale-high-gpu N=2       # 2 weithiwr blaenoriaeth uchel (keyboard + transcribe)
make scale-default-gpu N=3    # 3 gweithiwr rhagosodedig (transcribe_long_form)

# Gweithwyr CPU
make scale-high-cpu N=2
make scale-default-cpu N=2

Enghreifftiau Graddio

Canolbwyntio ar amser real (llawer o ddefnydd keyboard/transcribe):

make scale-high-gpu N=3       # Mwy o gapasiti blaenoriaeth uchel
make scale-default-gpu N=1    # Isafswm ffurf hir

Canolbwyntio ar brosesu swp (llawer o drawsgrifio ffurf hir):

make scale-high-gpu N=1       # Isafswm amser real
make scale-default-gpu N=3    # Mwy o gapasiti swp

Cytbwys (llwyth gwaith cymysg):

make scale-high-gpu N=2
make scale-default-gpu N=2

Ffurfweddiad Aml-GPU

Mae pob gweithiwr angen ~6GB VRAM. Mae'r ffurfweddiad rhagosodedig yn defnyddio dwy GPU:

GPU	Gweithwyr	VRAM a Ddefnyddir
GPU 1	`worker-high` + `worker-default`	~12GB
GPU 0	`worker-high-2`	~6GB

Mae hyn yn sicrhau bod ceisiadau blaenoriaeth uchel yn cael eu gwasanaethu'n gyflym hyd yn oed pan fo trawsgrifiad hir yn rhedeg ar GPU 1.

Ffurfweddu'r aseiniad GPU yn config.env:

CUDA_DEVICE_INDEX_HIGH=1      # worker-high → GPU 1
CUDA_DEVICE_INDEX_DEFAULT=1   # worker-default → GPU 1
CUDA_DEVICE_INDEX_HIGH_2=0    # worker-high-2 → GPU 0

Gweler docs/GPU_SCALING.md am fwy o fanylion.

Aseiniad Ciwiau Gweithwyr

Mae gweithwyr yn cael eu haseinio i giwiau trwy'r newidyn amgylchedd WORKER_QUEUES yn docker-compose:

# docker-compose.gpu.yml
services:
  worker-high:
    environment:
      - WORKER_QUEUES=high_priority    # Tasgau blaenoriaeth uchel (GPU 1)

  worker-high-2:
    environment:
      - WORKER_QUEUES=high_priority    # Tasgau blaenoriaeth uchel (GPU 0)

  worker-default:
    environment:
      - WORKER_QUEUES=default          # Tasgau rhagosodedig (GPU 1)

  worker-alignment:
    environment:
      - WORKER_QUEUES=alignment        # Tasgau alinio yn unig (CPU)

I greu gweithiwr sy'n trin y ddau giw (defnyddiol ar gyfer defnyddiadau bach):

WORKER_QUEUES=high_priority,default   # Yn gwirio high_priority yn gyntaf

Monitro

Gwiriadau Iechyd

# Iechyd cyffredinol y system
curl http://localhost:5511/health/

# Parodrwydd gweithwyr
curl http://localhost:5511/health/ready/

# Gwiriad bywiogrwydd
curl http://localhost:5511/health/live/

Statws Ciwiau

# Gwirio hyd ciwiau a thasgau gweithredol
curl http://localhost:5511/queue/status/

Logiau

# Gweld yr holl logiau
make logs

# Dilyn logiau gweithiwr blaenoriaeth uchel
docker compose -f docker-compose.gpu.yml logs -f worker-high

# Dilyn logiau gweithiwr rhagosodedig
docker compose -f docker-compose.gpu.yml logs -f worker-default

# Logiau fesul tasg (wedi'u storio yn recordings/<stt_id>.log)
curl http://localhost:5511/get_status/?stt_id=<uuid>

Gosod

Defnyddio Makefile (GPU yn ddiofyn)

git clone https://storfa.techiaith.cymru/lleferydd/stt-cy/whisperx-server
cd whisperx-server
make setup
make

Defnyddio Docker Compose yn Uniongyrchol

# Fersiwn GPU
docker-compose -f docker-compose.gpu.yml up -d --build

# Fersiwn CPU
docker-compose -f docker-compose.cpu.yml up -d --build

Yna, ewch wedyn i http://localhost:5511

Datrys Problemau

Modelau ddim yn llwytho

Gwiriwch y tocyn HuggingFace yn .env.secrets
Cadarnhewch gysylltedd rhwydwaith i HuggingFace
Adolygwch logiau gweithwyr: make logs

Gwallau allan o gof

Gostwng cydamseredd gweithwyr i 1
Galluogi ailgylchu max-tasks-per-child gweithwyr (eisoes wedi'i ffurfweddu)
Graddio i lawr nifer y gweithwyr
Defnyddio model Whisper llai

Trawsgrifio araf

Defnyddio modd GPU yn lle CPU
Gwirio a yw modelau wedi'u cadw (mae'r rhediad cyntaf yn llwytho modelau i lawr)
Graddio gweithwyr i fyny: make scale-high-gpu N=2 neu make scale-default-gpu N=3
Defnyddio /transcribe/ neu /keyboard/ ar gyfer tasgau brys (ciw blaenoriaeth uchel)

Tasgau'n sownd yn y ciw

Gwirio statws gweithwyr: make status
Archwilio'r ciw: curl http://localhost:5511/queue/status/
Ailgychwyn gweithwyr: docker compose -f docker-compose.gpu.yml restart worker-high worker-default worker-alignment

Datblygu

Strwythur y Prosiect

whisperx-server/
├── api/                    # Cymhwysiad FastAPI
│   ├── main.py            # Diweddbwyntiau API
│   ├── health.py          # Rhesymeg gwirio iechyd
│   ├── queue_inspector.py # Monitro ciwiau
│   └── cleanup_scheduler.py # Glanhau ffeiliau
├── worker/                 # Gweithiwr Celery
│   ├── tasks.py           # Tasg trawsgrifio
│   ├── speech_to_text_tasks.py # Llwytho model
│   ├── normalization/     # Normaleiddio testun Cymraeg
│   └── start_workers.sh   # Sgript cychwyn gweithwyr
├── shared/                 # Cyfleusterau a rennir
│   └── stt_logger.py      # Cyfleusterau logio
├── docs/                   # Dogfennaeth
├── config.env             # Ffurfweddiad tryloyw
└── .env.secrets.example   # Templed cyfrinachau

Trwydded

Trwydded MIT. Gweler LICENSE am fanylion.

Fersiwn Cymraeg uchod / Welsh version above

Key Benefits

🎯 Welsh Language Optimized

Fine-tuned models: Uses Welsh-specific Whisper and Wav2Vec2 models from Techiaith
Welsh normalization: Automatic text normalization for Welsh verbatim speech
- Expands contractions (ma'→mae, ca'l→cael)
- Standardizes dialectal variations (petha→pethau, gwbod→gwybod)
- Normalizes pronouns and verb forms
Language detection: Automatic Welsh/English detection with Lingua
Accurate alignment: Word-level timestamps optimized for Welsh phonetics

⚡ High Performance & Scalability

GPU acceleration: Full NVIDIA CUDA support for fast transcription
Dedicated queue workers: Separate high-priority and default workers, independently scalable
Task prioritization: Two-level priority system (queue routing + numeric priorities)
Efficient batching: WhisperX batched inference for 12x speedup
Memory management: Automatic cleanup and worker recycling to prevent memory leaks

🔧 Production Ready

Containerized deployment: Docker Compose for GPU and CPU environments
Health monitoring: Comprehensive health checks for all services
Automatic cleanup: Scheduled deletion of old files to prevent disk bloat
Task persistence: Redis-backed queue survives restarts
Logging: Per-task logging for debugging and monitoring
API documentation: Auto-generated OpenAPI/Swagger docs

📊 Rich Output Formats

JSON: Full transcription with word-level timestamps and scores
SRT/VTT: Subtitle files with intelligent line splitting
ELAN: Timeline annotations for linguistic analysis
Text: Plain text transcription
Speaker diarization: Identifies different speakers (SIARADWR_01, etc.)
Normalized text: Welsh-specific text normalization included

🚀 Developer Friendly

RESTful API: Simple HTTP endpoints for all operations
Queue inspection: Monitor pending tasks and worker status
Flexible deployment: CPU-only mode for development, GPU for production
Environment management: Transparent config with secret separation
Comprehensive docs: Architecture, deployment, and troubleshooting guides

🔒 Enterprise Features

Private model support: Authenticate with HuggingFace for private models
Task tracking: Persistent task IDs across server restarts
Result caching: 7-day result expiration with Redis backend
CORS enabled: Browser-friendly API access
Error handling: Graceful degradation and detailed error messages

Prerequisites

Docker and Docker Compose (required)
Linux or macOS — untested on Windows
GPU Mode: NVIDIA GPU + nvidia-docker2
Memory: Minimum 8GB RAM (16GB+ recommended for GPU mode)
Disk: At least 20GB free space for models and recordings

Quick Start

Initial Setup

# Clone repository
git clone https://storfa.techiaith.cymru/lleferydd/stt-cy/whisperx-server
cd whisperx-server

# Set up environment files
make setup

# Edit .env.secrets and add your HuggingFace token
# Get your token from: https://huggingface.co/settings/tokens
nano .env.secrets

# Regenerate .env with your token
make setup

GPU Deployment (Recommended for Production)

# Build and run with GPU support (requires nvidia-docker)
# Starts 1 high-priority worker + 1 default worker
make build-gpu

# Scale workers independently
make scale-high-gpu N=2       # 2 high-priority workers
make scale-default-gpu N=3    # 3 default workers

CPU Deployment (For Development/Testing)

# Build and run with CPU-only support (no GPU required)
# Starts 1 high-priority worker + 1 default worker
make build-cpu

# Scale workers independently
make scale-high-cpu N=2       # 2 high-priority workers
make scale-default-cpu N=2    # 2 default workers

Common Commands

make help          # Show all available commands
make ready         # Wait for models to finish loading
make status        # Show running containers
make health        # Check service health
make logs          # View container logs
make down          # Stop all containers

Access the API

Once running, access the API at: http://localhost:5511

Health Check: http://localhost:5511/health/ API Docs: http://localhost:5511/docs (FastAPI auto-generated)

Web App

A simple web app is available in the app/ folder for submitting transcriptions and translations via the browser. To use it:

python3 -m http.server 8080 --directory app/

Then open http://localhost:8080 in your browser. The app connects to the API on localhost:5511.

API Usage Examples

Transcribe Audio (Short Form, Synchronous)

# Default: high priority (routed to high_priority queue)
curl -X POST "http://localhost:5511/transcribe/" \
  -F "soundfile=@audio.wav"

# Explicit normal priority (still routed to high_priority queue)
curl -X POST "http://localhost:5511/transcribe/" \
  -F "soundfile=@audio.wav" \
  -F "priority=normal"

Transcribe Long Audio (Asynchronous)

# Submit task (routed to default queue)
curl -X POST "http://localhost:5511/transcribe_long_form/" \
  -F "soundfile=@long_audio.wav"

# Returns: {"id": "uuid", "version": 2, "success": true}

# Check status
curl "http://localhost:5511/get_status/?stt_id=<uuid>"

# Get results when complete
curl "http://localhost:5511/get_json/?stt_id=<uuid>"
curl "http://localhost:5511/get_srt/?stt_id=<uuid>"
curl "http://localhost:5511/get_vtt/?stt_id=<uuid>"

Translate Audio (Welsh → English)

# Short form (synchronous, max 480KB)
curl -X POST "http://localhost:5511/translate/" \
  -F "soundfile=@audio.wav"

# Long form (asynchronous, any size)
curl -X POST "http://localhost:5511/translate_long_form/" \
  -F "soundfile=@long_audio.wav"

Align Text to Audio (Word-level Timestamps)

# Short form (synchronous, max 480KB)
curl -X POST "http://localhost:5511/align/" \
  -F "soundfile=@audio.wav" \
  -F "text=Mae ganddynt dau o blant, mab a merch"

# Long form (asynchronous, any size — runs on CPU)
curl -X POST "http://localhost:5511/align_long_form/" \
  -F "soundfile=@long_audio.wav" \
  -F "text=Full text to align against the audio..."

# Check status
curl "http://localhost:5511/get_status/?stt_id=<uuid>"

# Get results
curl "http://localhost:5511/get_json/?stt_id=<uuid>"

Real-time Keyboard Input (Always High Priority)

curl -X POST "http://localhost:5511/keyboard/" \
  -F "audio_file=@voice_input.wav"
# Always routed to high_priority queue with urgent priority

Architecture

Components

┌──────────────────┐
│    FastAPI App    │  Port 5511
│   (API Server)   │
└────────┬─────────┘
         │
         ▼
┌──────────────────┐
│   Redis Broker   │  Port 6379
│   (Task Queue)   │
├──────────────────┤
│ high_priority    │  /transcribe/, /keyboard/, /translate/
│ default          │  /transcribe_long_form/, /translate_long_form/
│ alignment        │  /align/, /align_long_form/
└──┬───┬───┬───┬───┘
   │   │   │   │
   ▼   ▼   ▼   ▼
┌────────┐ ┌──────────┐ ┌─────────┐ ┌───────────┐
│worker  │ │worker    │ │worker   │ │worker     │  Independently
│-high   │ │-high-2   │ │-default │ │-alignment │  Scalable
│GPU 1   │ │GPU 0     │ │GPU 1    │ │(CPU x2)   │
│(GPU 1) │ │(GPU 0)   │ │(GPU 1)  │ │           │
└────────┘ └──────────┘ └─────────┘ └───────────┘

GPU Assignment

Service	Queue	GPU	Notes
`worker-high`	`high_priority`	GPU 1	Shares GPU with worker-default
`worker-high-2`	`high_priority`	GPU 0	Dedicated GPU, high-priority overflow
`worker-default`	`default`	GPU 1	Shares GPU with worker-high
`worker-alignment` (x2)	`alignment`	CPU	wav2vec2 alignment on CPU

When worker-default is busy with a long-form transcription, worker-high-2 on GPU 0 ensures high-priority requests are still served quickly.

Services

Application (FastAPI): REST API server handling requests
worker-high (Celery): Dedicated to high-priority tasks (/transcribe/, /keyboard/, /translate/) — GPU 1
worker-high-2 (Celery): Second high-priority worker on dedicated GPU (/transcribe/, /keyboard/, /translate/) — GPU 0
worker-default (Celery): Dedicated to default tasks (/transcribe_long_form/, /translate_long_form/) — GPU 1
worker-alignment (Celery): Dedicated CPU alignment worker (/align/, /align_long_form/) — 2 replicas so a long task doesn't block short requests
Redis: Message broker and result backend

Task Queue & Priority System

The system uses two-level prioritization: dedicated queue workers + numeric priorities within each queue.

Queue Routing by Endpoint

Endpoint	Queue	Default Priority	Notes
`/keyboard/`	`high_priority`	0 (urgent)	Always urgent, no priority parameter
`/transcribe/`	`high_priority`	0 (high)	Accepts `high` or `normal`
`/translate/`	`high_priority`	0 (high)	Welsh → English translation
`/transcribe_long_form/`	`default`	5 (normal)	Asynchronous, no priority parameter
`/translate_long_form/`	`default`	5 (normal)	Async translation, no priority parameter
`/align/`	`alignment`	0 (urgent)	Align text to audio, word-level timestamps (CPU)
`/align_long_form/`	`alignment`	5 (normal)	Async long audio alignment (CPU)

Worker Queue Assignment

Workers are dedicated to specific queues via the WORKER_QUEUES environment variable:

Worker	Queues	Device	Serves
`worker-high`	`high_priority`	GPU 1	`/transcribe/`, `/keyboard/`, `/translate/`
`worker-high-2`	`high_priority`	GPU 0	`/transcribe/`, `/keyboard/`, `/translate/`
`worker-default`	`default`	GPU 1	`/transcribe_long_form/`, `/translate_long_form/`
`worker-alignment` (x2)	`alignment`	CPU	`/align/`, `/align_long_form/`

Numeric Priorities (within each queue)

Value	Name	Used By
0	urgent	`/keyboard/`, `/transcribe/` with `priority=high`, `/translate/`, `/align/`
5	normal	`/transcribe/` with `priority=normal`, `/transcribe_long_form/`, `/translate_long_form/`, `/align_long_form/`

Output Format

JSON Response

{
  "id": "audio_id",
  "version": 2,
  "success": true,
  "language": "cy",
  "segments": [
    {
      "audio_id": "md5_hash",
      "start": 0.0,
      "end": 2.5,
      "text": "Mae hyn yn brawf o drawsgrifio",
      "normalized": "Mae hyn yn brawf o drawsgrifio",
      "score": 0.95,
      "words": [
        {"word": "Mae", "start": 0.0, "end": 0.3, "score": 0.98},
        {"word": "hyn", "start": 0.3, "end": 0.5, "score": 0.96}
      ],
      "chars": [...]
    }
  ]
}

Features

Text normalization: Each segment includes both original text and normalized (standardized Welsh)
Word-level timestamps: Precise timing for each word
Confidence scores: Per-word and per-segment accuracy scores
Speaker diarization: Speaker labels (SIARADWR_01, SIARADWR_02, etc.)
Language detection: Automatic Welsh/English identification

Environment Configuration

This project uses a hybrid approach for environment variables:

File Structure

config.env - Transparent configuration (committed to git)
- Model names, Redis settings, non-sensitive defaults
- Edit this file to change model configurations
.env.secrets - Sensitive values (gitignored)
- HuggingFace tokens, passwords
- Create from .env.secrets.example
.env - Generated file (gitignored)
- Automatically created by make setup
- Merges config.env + .env.secrets

Key Configuration Options

# Model Configuration (config.env)
WHISPER_MODEL_NAME=DewiBrynJones/whisper-large-v2-ft-btb-cv-cvad-ca-wlga-cy-ct2-2511
WHISPER_MODEL_LANGUAGE=cy
WAV2VEC2_MODEL=techiaith/wav2vec2-btb-cv-ft-cv-cy

# File Cleanup (config.env)
CLEANUP_ENABLED=true
FILE_RETENTION_DAYS=14
CLEANUP_SCHEDULE=0 2 * * *  # 2 AM daily

# Secrets (.env.secrets)
HF_AUTH_TOKEN=your_huggingface_token_here

Modifying Configuration

To change model settings:

# Edit the transparent config file
nano config.env

# Regenerate .env
make setup

To update your HuggingFace token:

# Edit secrets file
nano .env.secrets

# Regenerate .env
make setup

Documentation

Core Documentation

CPU_GPU_DEPLOYMENT.md - Complete guide for CPU vs GPU deployment
CELERY_ARCHITECTURE.md - Understanding the task queue system
DOCKER_HEALTHCHECKS.md - Container health monitoring

Feature Documentation

FILE_CLEANUP.md - Automatic cleanup of old transcription files
TASK_PRIORITIZATION.md - Queue routing and priority system
IMPROVING_ALIGNMENT.md - Enhancing word-level timestamp accuracy

Scaling & Deployment

GPU_SCALING.md - GPU memory limits and multi-GPU configurations
ADVANCED_SCALING.md - Advanced scaling strategies
INFERENCE_DEPLOYMENT_ALTERNATIVES.md - Comparison of inference deployment solutions

macOS Native Deployment

macos/README.md - Native Apple Silicon deployment with MLX acceleration
macos/HYBRID_MLX_ARCHITECTURE.md - MLX hybrid architecture documentation

Technical Documentation

claude_refactors_2601.md - Recent architectural improvements

Scaling

Independent Worker Scaling

Scale high-priority and default workers independently based on your workload:

# GPU workers
make scale-high-gpu N=2       # 2 high-priority workers (keyboard + transcribe)
make scale-default-gpu N=3    # 3 default workers (transcribe_long_form)

# CPU workers
make scale-high-cpu N=2
make scale-default-cpu N=2

Scaling Examples

Real-time focused (lots of keyboard/transcribe usage):

make scale-high-gpu N=3       # More high-priority capacity
make scale-default-gpu N=1    # Minimal long-form

Batch processing focused (lots of long-form transcription):

make scale-high-gpu N=1       # Minimal real-time
make scale-default-gpu N=3    # More batch capacity

Balanced (mixed workload):

make scale-high-gpu N=2
make scale-default-gpu N=2

Multi-GPU Configuration

Each worker needs ~6GB VRAM. The default configuration uses two GPUs:

GPU	Workers	VRAM Used
GPU 1	`worker-high` + `worker-default`	~12GB
GPU 0	`worker-high-2`	~6GB

This ensures high-priority requests are served quickly even when a long-form transcription is running on GPU 1.

Configure GPU assignment in config.env:

CUDA_DEVICE_INDEX_HIGH=1      # worker-high → GPU 1
CUDA_DEVICE_INDEX_DEFAULT=1   # worker-default → GPU 1
CUDA_DEVICE_INDEX_HIGH_2=0    # worker-high-2 → GPU 0

See docs/GPU_SCALING.md for more details.

Worker Queue Assignment

Workers are assigned to queues via the WORKER_QUEUES environment variable in docker-compose:

# docker-compose.gpu.yml
services:
  worker-high:
    environment:
      - WORKER_QUEUES=high_priority    # High-priority tasks (GPU 1)

  worker-high-2:
    environment:
      - WORKER_QUEUES=high_priority    # High-priority tasks (GPU 0)

  worker-default:
    environment:
      - WORKER_QUEUES=default          # Default tasks (GPU 1)

  worker-alignment:
    environment:
      - WORKER_QUEUES=alignment        # Alignment tasks only (CPU)

To create a worker that handles both queues (useful for small deployments):

WORKER_QUEUES=high_priority,default   # Checks high_priority first

Monitoring

Health Checks

# Overall system health
curl http://localhost:5511/health/

# Worker readiness
curl http://localhost:5511/health/ready/

# Liveness check
curl http://localhost:5511/health/live/

Queue Status

# Check queue lengths and active tasks
curl http://localhost:5511/queue/status/

Logs

# View all logs
make logs

# Follow high-priority worker logs
docker compose -f docker-compose.gpu.yml logs -f worker-high

# Follow default worker logs
docker compose -f docker-compose.gpu.yml logs -f worker-default

# Per-task logs (stored in recordings/<stt_id>.log)
curl http://localhost:5511/get_status/?stt_id=<uuid>

Installation

Using Makefile (GPU by default)

git clone https://storfa.techiaith.cymru/lleferydd/stt-cy/whisperx-server
cd whisperx-server
make setup
make

Using Docker Compose Directly

# GPU version
docker-compose -f docker-compose.gpu.yml up -d --build

# CPU version
docker-compose -f docker-compose.cpu.yml up -d --build

Then navigate to http://localhost:5511

Troubleshooting

Models not loading

Check HuggingFace token in .env.secrets
Verify network connectivity to HuggingFace
Review worker logs: make logs

Out of memory errors

Reduce worker concurrency to 1
Enable max-tasks-per-child worker recycling (already configured)
Scale down number of workers
Use smaller Whisper model

Slow transcription

Use GPU mode instead of CPU
Check if models are cached (first run downloads models)
Scale up workers: make scale-high-gpu N=2 or make scale-default-gpu N=3
Use /transcribe/ or /keyboard/ for urgent tasks (high-priority queue)

Tasks stuck in queue

Check worker status: make status
Inspect queue: curl http://localhost:5511/queue/status/
Restart workers: docker compose -f docker-compose.gpu.yml restart worker-high worker-default worker-alignment

Development

Project Structure

whisperx-server/
├── api/                    # FastAPI application
│   ├── main.py            # API endpoints
│   ├── health.py          # Health check logic
│   ├── queue_inspector.py # Queue monitoring
│   └── cleanup_scheduler.py # File cleanup
├── worker/                 # Celery worker
│   ├── tasks.py           # Transcription task
│   ├── speech_to_text_tasks.py # Model loading
│   ├── normalization/     # Welsh text normalization
│   └── start_workers.sh   # Worker startup script
├── shared/                 # Shared utilities
│   └── stt_logger.py      # Logging utilities
├── docs/                   # Documentation
├── config.env             # Transparent configuration
└── .env.secrets.example   # Secret template

License

MIT License. See LICENSE for details.

Name		Name	Last commit message	Last commit date
Latest commit History 6 Commits
api		api
app		app
shared		shared
worker		worker
.dockerignore		.dockerignore
.env.secrets.example		.env.secrets.example
.gitignore		.gitignore
.gitmodules		.gitmodules
CLAUDE.md		CLAUDE.md
Dockerfile.cpu		Dockerfile.cpu
Dockerfile.gpu		Dockerfile.gpu
LICENSE		LICENSE
Makefile		Makefile
README.md		README.md
config.env		config.env
copy-shared.sh		copy-shared.sh
docker-compose.cpu.yml		docker-compose.cpu.yml
docker-compose.gpu.yml		docker-compose.gpu.yml
speech.wav		speech.wav

Folders and files

Latest commit

History

Repository files navigation

welsh-whisperx

Prif Fanteision

🎯 Wedi'i Optimeiddio ar gyfer y Gymraeg

⚡ Perfformiad Uchel a Graddadwyedd

🔧 Yn Barod ar gyfer Cynhyrchu

📊 Fformatau Allbwn Cyfoethog

🚀 Cyfeillgar i Ddatblygwyr

🔒 Nodweddion Menter

Prif Fanteision

🎯 Wedi'i Optimeiddio ar gyfer y Gymraeg

⚡ Perfformiad Uchel a Graddadwyedd

🔧 Yn Barod ar gyfer Cynhyrchu

📊 Fformatau Allbwn Cyfoethog

🚀 Cyfeillgar i Ddatblygwyr

🔒 Nodweddion Menter

Rhagofynion

Cychwyn Cyflym

Gosod Cychwynnol

Defnyddio GPU (Argymhellir ar gyfer Cynhyrchu)

Defnyddio CPU (Ar gyfer Datblygu/Profi)

Gorchmynion Cyffredin

Mynediad i'r API

Ap Gwe

Enghreifftiau Defnyddio'r API

Trawsgrifio Sain (Ffurf Fer, Cydamserol)

Trawsgrifio Sain Hir (Anghydamserol)

Cyfieithu Sain (Cymraeg → Saesneg)

Alinio Testun â Sain (Stampiau Amser fesul Gair)

Mewnbwn Bysellfwrdd Amser Real (Blaenoriaeth Uchel Bob Amser)

Pensaernïaeth

Cydrannau

Aseiniad GPU

Gwasanaethau

System Ciwiau a Blaenoriaethu Tasgau

Llwybro Ciwiau fesul Diweddbwynt

Aseiniad Ciwiau Gweithwyr

Blaenoriaethau Rhifol (o fewn pob ciw)

Fformat Allbwn

Ymateb JSON

Nodweddion

Ffurfweddiad yr Amgylchedd

Strwythur Ffeiliau

Prif Opsiynau Ffurfweddu

Newid Ffurfweddiad

Dogfennaeth

Dogfennaeth Graidd

Dogfennaeth Nodweddion

Graddio a Defnyddio

Defnyddio Brodorol macOS

Dogfennaeth Dechnegol

Graddio

Graddio Gweithwyr yn Annibynnol

Enghreifftiau Graddio

Ffurfweddiad Aml-GPU

Aseiniad Ciwiau Gweithwyr

Monitro

Gwiriadau Iechyd

Statws Ciwiau

Logiau

Gosod

Defnyddio Makefile (GPU yn ddiofyn)

Defnyddio Docker Compose yn Uniongyrchol

Datrys Problemau

Modelau ddim yn llwytho

Gwallau allan o gof

Trawsgrifio araf

Tasgau'n sownd yn y ciw

Datblygu

Strwythur y Prosiect

Trwydded

Key Benefits

🎯 Welsh Language Optimized

⚡ High Performance & Scalability

🔧 Production Ready

📊 Rich Output Formats

🚀 Developer Friendly