# Note: All the variables to this image are optional.
# See README for more information.
# To use, copy this file to 'whisper.env' and uncomment/set your values.

# Whisper model to use for transcription (default: base)
# Options: tiny, tiny.en, base, base.en, small, small.en, medium, medium.en,
#          large-v1, large-v2, large-v3, large-v3-turbo, turbo
# - Models are downloaded from HuggingFace on first start and cached in the
#   /var/lib/whisper Docker volume.
# - large-v3-turbo is recommended if you need better accuracy than 'base'
#   without the full resource cost of large-v3.
# - Use an '.en' variant (e.g. base.en) for a slight speed boost on English-only audio.
# WHISPER_MODEL=base

# Default language for transcription (default: auto)
# - Set to a BCP-47 language code (e.g. en, fr, de, zh, ja, es) to skip
#   language autodetection and improve speed/accuracy for known-language audio.
# - Leave as 'auto' or unset to autodetect the language from each audio file.
# WHISPER_LANGUAGE=auto

# HTTP port for the transcription API (default: 9000)
# WHISPER_PORT=9000

# Compute device (default: cpu)
# Options: cpu, cuda, auto
# - cpu:   Use CPU for inference (default, works everywhere)
# - cuda:  Use NVIDIA GPU (requires the :cuda image tag and NVIDIA Container Toolkit)
# - auto:  Automatically detect GPU; falls back to CPU if unavailable
# WHISPER_DEVICE=cpu

# Quantization / compute type (default: int8 on CPU, float16 on CUDA)
# - int8          Fastest on CPU; good accuracy. Recommended for CPU deployments.
# - float16       Recommended for GPU (CUDA). Fast and accurate on NVIDIA GPUs.
# - int8_float16  Mixed precision; useful on newer CPUs with AVX512 or on GPU.
# - float32       Highest precision; slowest. Rarely needed.
# WHISPER_COMPUTE_TYPE=int8

# Number of CPU threads used for inference (default: 2)
# - Increase on hosts with many cores to improve throughput for concurrent requests.
# - Set to the number of physical CPU cores for best single-request latency.
# WHISPER_THREADS=2

# Optional API key to require Bearer token authentication on all requests.
# - Fresh installs with a mounted /var/lib/whisper volume auto-generate a key.
# - Existing installs with no key remain open for backward compatibility.
# - When set, every request must include: -H "Authorization: Bearer <key>"
# - Set explicitly empty (WHISPER_API_KEY=) to disable authentication.
# - Use 'docker exec <container> whisper_manage --showkey' to display the key.
# - Use 'docker exec <container> whisper_manage --getkey' for scripts.
# - Keep this secret if you expose the server to the internet.
# WHISPER_API_KEY=

# Log level (default: INFO)
# Options: DEBUG, INFO, WARNING, ERROR, CRITICAL
# WHISPER_LOG_LEVEL=INFO

# Beam size for transcription and translation decoding (default: 5)
# - Controls how many candidate sequences are explored simultaneously during
#   decoding. Higher values can improve accuracy at the cost of more CPU/memory.
# - Use 1 for fastest (greedy) decoding; 5 is a good accuracy/speed balance.
# WHISPER_BEAM=5

# Maximum beam size allowed for the per-request beam override (default: 10)
# - Requests may set beam=N as a local faster-whisper extension.
# - Set to 0 to allow any positive request beam value.
# WHISPER_MAX_REQUEST_BEAM=10

# Maximum uploaded audio file size in MB (default: 1024)
# - Requests larger than this limit are rejected with HTTP 413.
# - Set to 0 to disable the upload size limit.
# WHISPER_MAX_UPLOAD_MB=1024

# Local-only mode: disable all HuggingFace model downloads (default: unset)
# - When set to any non-empty value (e.g. true), faster-whisper will only use
#   models already present in the /var/lib/whisper cache volume and will not
#   attempt to contact HuggingFace. Useful for air-gapped or offline deployments.
# - Leave unset (default) to allow automatic model downloads on first start.
# WHISPER_LOCAL_ONLY=

# Enable word-level timestamps globally (default: unset / false)
# - When set to 'true', all transcription requests will include word-level
#   timestamps in the verbose_json output (a 'words' array in each segment
#   with per-word start/end times and confidence scores).
# - Can also be enabled per-request by passing word_timestamps=true in the
#   API call, without setting this variable.
# WHISPER_WORD_TIMESTAMPS=

# --- Speaker Diarization (optional) ---
# Enable speaker diarization to identify who is speaking in each segment.
# Uses sherpa-onnx with pyannote segmentation-3.0 ONNX models.
# ONNX models (~45 MB total) are auto-downloaded on first use.
# Diarization is not supported in streaming mode (silently skipped).
# WHISPER_DIARIZATION=true

# Exact number of speakers (if known). Improves clustering accuracy.
# Leave unset or -1 for automatic detection.
# WHISPER_DIARIZE_NUM_SPEAKERS=

# Clustering threshold for automatic detection (default: 0.5).
# Lower = more speakers, higher = fewer. Ignored when exact speaker count is set.
# WHISPER_DIARIZE_THRESHOLD=0.5

# Anonymous aggregate usage counts (default: enabled)
# Set to 1 to disable.
# WHISPER_DISABLE_USAGE_COUNTS=1