speakerscribe 0.1.1

Speech-to-text with speaker diarization — Whisper + pyannote.audio, optimized for Google Colab Free Tier (T4 GPU).

speakerscribe

Speech-to-text with speaker diarization — powered by faster-whisper + pyannote.audio, optimized for Google Colab Free Tier (T4 GPU).

---

What it does

speakerscribe takes any audio or video file and produces:

Output	Description
.txt	Plain transcript with [SPEAKER_XX] labels per line
.srt	Subtitle file with timestamps and speaker labels
.json	Full structured metadata (segments, speakers, versions, RTF, timings)
.transcript.md	Readable Markdown grouped by speaker turns with a metadata header
_1.txt, _2.txt …	Text chunks of ~1950 words for downstream LLM processing
_full_for_llm.txt	Single-file full transcript for LLMs with large context windows
.segments.jsonl	Per-segment streaming output (opt-in via streaming_jsonl=True)

---

Quick start (Google Colab — recommended)

Open the companion notebook:

The notebook handles everything: installation, Google Drive mounting, HF token setup, and batch processing.

---

Installation

pip install speakerscribe

Requires Python ≥ 3.10. On Colab, restart the runtime after installation.

HuggingFace token (required for diarization)

pyannote.audio requires a free HuggingFace token with access to the diarization model:

1. Create a Read token at https://huggingface.co/settings/tokens (do not use fine-grained tokens).
2. Accept the model terms at https://huggingface.co/pyannote/speaker-diarization-community-1.
3. Make the token available in one of three ways:
   • Colab Secrets (recommended): add HF_TOKEN in the Colab sidebar.
   • Environment variable: export HF_TOKEN=hf_...
   • Config parameter: TranscriptionConfig(hf_token="hf_...")

---

Python API

from speakerscribe import TranscriptionConfig, WorkspacePaths, process_batch

# 1. Configure the pipeline
config = TranscriptionConfig(
    model="large-v3-turbo",   # recommended: fast + accurate
    language="es",            # None = auto-detect
    beam_size=5,
    enable_diarization=True,
    # hf_token="hf_..."       # or set HF_TOKEN env var / Colab Secret
)

# 2. Set workspace (put your audio files in <workspace>/data/)
paths = WorkspacePaths(workspace="/content/drive/MyDrive/MyProject")

# 3. Run — processes all files in data/ and writes outputs to transcripts/
results = process_batch(paths, config)

Process a single file

from speakerscribe import TranscriptionConfig, WorkspacePaths
from speakerscribe.pipeline import preflight_check, process_one
from speakerscribe.transcription import load_whisper_model, release_whisper_model

config = TranscriptionConfig(model="large-v3-turbo", language="en")
paths  = WorkspacePaths(workspace="/tmp/my_project")

preflight_check(paths, config)
model = load_whisper_model(config)
try:
    result = process_one(paths.data / "interview.mp4", paths, model, config)
    print(result["total_words"], "words transcribed")
finally:
    release_whisper_model(model)

Rename SPEAKER_XX labels after reviewing the transcript

from speakerscribe import WorkspacePaths, rename_speakers_in_outputs

paths = WorkspacePaths(workspace="/tmp/my_project")
rename_speakers_in_outputs(
    paths,
    base_name="interview_large-v3-turbo",
    mapping={"SPEAKER_00": "Alice", "SPEAKER_01": "Bob"},
)

---

Command-line interface

# Process all audio files in a workspace
speakerscribe process --workspace /path/to/project --model large-v3-turbo --language es

# Enable debug logging
speakerscribe --verbose process --workspace /path/to/project

# Load config from a JSON file instead of CLI flags
speakerscribe process --workspace /path/to/project --config-file config.json

# Run a fast smoke test (small model, first file only)
speakerscribe smoke-test --workspace /path/to/project

# Inspect a transcription JSON file
speakerscribe inspect /path/to/project/transcripts/interview.json

# Show aggregate statistics from the runs database
speakerscribe stats --workspace /path/to/project

# List the most recent runs
speakerscribe list-runs --workspace /path/to/project --limit 20

# Rename speaker labels
speakerscribe rename --workspace /path/to/project \
  --base-name "interview_large-v3-turbo" \
  --mapping mapping.json

# Delete outputs for a specific file (keep source audio)
speakerscribe clean --workspace /path/to/project --pattern "interview"

# Delete everything (requires explicit confirmation)
speakerscribe clean --workspace /path/to/project --all --confirm "YES DELETE ALL"

# Show version
speakerscribe version

---

Workspace layout

my_project/
├── data/                  ← place your .mp4 / .mp3 / .wav / .m4a files here
├── transcripts/           ← outputs: .txt, .srt, .json, .transcript.md
├── splits/                ← chunked .txt files and _full_for_llm.txt for LLM use
├── _audio_temp/           ← temporary 16 kHz WAVs (auto-deleted)
├── _audio_chunks/         ← chunked WAVs for long-audio splitting (auto-deleted)
├── _diar_cache/           ← cached pyannote diarization results (reuse on reruns)
├── _logs/                 ← rotating log files
└── _runs.db               ← SQLite run history (created when enable_runs_db=True)

---

Supported models

Model	VRAM (T4)	Speed	Quality
tiny	~1.5 GB	fastest	basic
base	~1.8 GB	very fast	fair
small	~2.5 GB	fast	good
medium	~3.5 GB	moderate	very good
large-v3-turbo ⭐	~3.5 GB	fast	excellent
large-v3	~5.0 GB	slower	best

large-v3-turbo is the recommended choice for most workflows: ~3× faster than large-v3 with comparable accuracy. The default when no model is specified is large-v3.

---

Supported audio/video formats

.mp4 · .mp3 · .wav · .m4a · .mkv · .aac · .flac · .ogg · .webm

Audio is automatically converted to 16 kHz mono WAV via ffmpeg before processing.

---

Configuration reference

TranscriptionConfig(
    # ── Model ──────────────────────────────────────────────────────
    model            = "large-v3",        # Whisper model name (default: large-v3)
    device           = "auto",            # "auto" | "cuda" | "cpu"
    compute_type     = "auto",            # "auto" | "float16" | "int8"

    # ── Decoding ───────────────────────────────────────────────────
    beam_size        = 5,                 # 1 (greedy) to 10 (highest quality)
    language         = None,              # None = auto-detect; "en", "es", etc.
    initial_prompt   = None,              # glossary string for proper nouns / jargon
                                          # per-file override: <stem>.prompt.txt next to audio

    # ── VAD ────────────────────────────────────────────────────────
    use_vad              = True,          # Silero VAD — skip silences
    vad_min_silence_ms   = 500,           # silences ≥ N ms split segments

    # ── Transcription ──────────────────────────────────────────────
    word_timestamps  = False,             # per-word timestamps (slower)

    # ── Diarization ────────────────────────────────────────────────
    enable_diarization   = True,          # set False for transcription only
    hf_token             = None,          # or set HF_TOKEN env var / Colab Secret
    num_speakers         = None,          # pin exact count if known
    min_speakers         = None,          # or use min/max range
    max_speakers         = None,
    diarization_model    = "pyannote/speaker-diarization-community-1",

    # ── Transcript output ──────────────────────────────────────────
    gap_max_s_transcript    = 2.0,        # seconds of silence to open a new speaker turn
    generate_transcript_md  = True,       # write the readable .transcript.md
    remove_fillers          = True,       # drop filler-only segments from .transcript.md
                                          # (.json and .srt keep ALL segments)

    # ── Splits and LLM output ──────────────────────────────────────
    words_per_split          = 1950,      # chunk size for split files
    produce_unified_for_llm  = True,      # also write a single _full_for_llm.txt

    # ── Long-audio chunking ────────────────────────────────────────
    long_audio_threshold_min = 120,       # audios longer than N min are chunked
    chunk_duration_min       = 30,        # target chunk length in minutes
    chunk_overlap_s          = 5,         # overlap between chunks in seconds

    # ── Streaming ─────────────────────────────────────────────────
    streaming_jsonl  = False,             # write segments to .segments.jsonl as produced

    # ── Run control ────────────────────────────────────────────────
    force_reprocess  = False,             # True = ignore existing outputs
    evaluate_quality = True,              # heuristic quality check after each file
    enable_runs_db   = True,              # SQLite run history for idempotency by hash

    # ── Disk safety ────────────────────────────────────────────────
    disk_margin_factor  = 0.5,            # required free space = input_size × factor
    disk_margin_min_mb  = 500,            # hard floor for required free space (MB)
)

---

Quality checker

After each file, speakerscribe runs a heuristic quality check and logs any issues:

Flag	Severity	Meaning
LOW_LANG_CONFIDENCE	WARNING	Language detection < 85% — check if audio is noisy
LOW_RTF	WARNING	Processing faster than 2× — possible pipeline issue
HIGH_RTF	INFO	Audio mostly silence
LOW_WPM	WARNING	< 60 words/min — aggressive VAD or very slow speech
HIGH_WPM	CRITICAL	> 250 words/min — likely Whisper hallucination
SPEAKER_DOMINANCE	WARNING	One speaker > 95% — poor diarization
TOO_MANY_SPEAKERS	WARNING	> 8 speakers detected — consider pinning num_speakers
TINY_SPEAKERS	WARNING	Speakers with ≤ 2 segments — likely false positives
REPETITIONS	CRITICAL	Consecutive repeated n-grams — Whisper hallucination loop
WORD_DOMINANCE	WARNING	One word dominates the transcript — possible hallucination
EMPTY_SEGMENTS	WARNING	> 10% empty segments — VAD too aggressive

---

Requirements

• Python ≥ 3.10
• ffmpeg installed and available in PATH
• NVIDIA GPU (T4 or better) recommended; CPU mode is also supported but slow

On Google Colab

ffmpeg is pre-installed. GPU runtime: Runtime → Change runtime type → T4 GPU.

---

Development

git clone https://github.com/EnriqueForero/speakerscribe
cd speakerscribe
pip install -e ".[dev]"
pre-commit install

# Run tests (no GPU needed for unit tests)
pytest tests/ -m "not integration and not gpu"

# Lint
ruff check speakerscribe/ tests/
ruff format speakerscribe/ tests/

---

License

MIT — © 2026 Néstor Enrique Forero Herrera