Parlando (Italian: “speaking” — musical and theatrical recitation delivered with the natural cadence of speech) is a studio-grade neural prose synthesis and audio engineering engine.
Designed specifically for fiction writers, long-form manuscripts, and online articles, Parlando transforms raw prose into fluid, broadcast-mastered audiobooks with generic multi-character speaker detection, scene-aware voice casting, intelligent punctuation pacing, zero-crossing DSP stitching, and interactive Web Audio players.
Parlando operates as a standalone CLI tool, a containerized REST microservice, and the official speech synthesis companion for Marginalia.
┌────────────────────────┐
│ Manuscript Ingest │
│ (TXT, MD, EPUB, URLs) │
└───────────┬────────────┘
│
┌───────────▼────────────┐
│ Narrative Chunker │
│ (Dialogue & Headings) │
└───────────┬────────────┘
│
┌───────────▼────────────┐
│ Generic Speaker Engine │
│(Attribution & Pronouns)│
└───────────┬────────────┘
│
┌───────────▼────────────┐
│ Scene-Aware Voice Cast │
│ (Graph-Color Allocator)│
└───────────┬────────────┘
│
┌───────────▼────────────┐
│ Prosody Director │
│ (Pacing & Envelopes) │
└───────────┬────────────┘
│
┌───────────▼────────────┐
│ Pluggable Neural TTS │
│ (Edge / OAI / Gemini) │
└───────────┬────────────┘
│
┌───────────▼────────────┐
│ Audio DSP & Stitcher │
│ (Zero-Crossing + Fade) │
└───────────┬────────────┘
│
┌───────────▼────────────┐
│ EBU R128 Mastering │
│ (M4B / MP3 / Chapters) │
└───────────┬────────────┘
│
┌───────────▼────────────┐
│ Standalone Web Player │
│ (HTML5 + FFT Visual) │
└────────────────────────┘
Experience Parlando's neural prosody, multi-character dialogue isolation, and zero-crossing DSP in action synthesizing William Gibson's Chiba City Blues (Neuromancer Chapters 1 & 2):
▶ Play / Download MP3 (2.10 min)- 🔊 Direct Audio Stream: Play
chiba_city_blues.mp3(2.10 min) (19 narrative chunks, multi-speaker Gemini neural audio attribution:FenrirNarrator +PuckCase +AoedeLinda Lee +CharonHotel Clerk) - 🌐 Interactive Web Player: Launch Web Player Online or open
samples/chiba_city_blues_player.html(Features chapter scrubbing, speed controls, and real-time spectrum visualizer with 100% self-contained embedded audio) - 📄 Source Manuscript:
samples/neuromancer_sample.md
parlando samples/neuromancer_sample.md \
-o samples/chiba_city_blues.mp3 \
--format mp3 \
--backend gemini \
--voice Fenrir \
-c "Case:male:Puck" \
-c "Linda Lee:female:Aoede" \
--pacing dramatic(With character voice attributions: Narrator -> Fenrir, Case -> Puck, Linda Lee -> Aoede, Hotel Clerk -> Charon)
- Generic Manuscript Parsing: Automatically identifies speaking characters across any arbitrary story or genre without hardcoded rules.
- Deterministic Pronoun & Verb Attribution: Resolves pronouns (
she,he,they) and syntactic dialogue tags (said, whispered, rasped, murmured, demanded) to specific character entities. - Consistent Narrator Invariance: The primary narrator voice (
--voice) is strictly reserved for exposition, chapter headings, and section breaks. - Scene-Aware Graph Coloring: Characters interacting in the same conversation scene are guaranteed distinct voices. Characters who appear in separate scenes can safely reuse voices from the pool without listener confusion.
- More Characters Than Voices:
- Interaction Graph Multiplexing: A 30-character novel only requires as many voices as the largest single conversation scene (typically 2–4).
- Line-Count Prioritization: Major recurring characters receive first pick of distinct primary voices.
- Cross-Pool Fallbacks: Dynamically draws from neutral and expanded voice pools.
- Prosody Modulation: If dialogue reuse is necessary within a dense scene, Parlando applies pitch (
±2Hz) and pacing (±4%) offsets to render acoustically distinct personas.
- CLI Custom Cast: Specify character names, genders, and optional voice overrides via repeatable
-c / --characterflags or--cast cast.json. - Flexible Programmatic API: Pass dictionaries, dataclasses, or lists into
PipelineConfig(characters=...). - Smart Auto-Merge: User-specified characters and voices are locked in, while any remaining manuscript characters are automatically detected and cast.
- Contraction-Safe Segmentation: Preserves word-internal apostrophes (
don't,Manfred's,we've,I'm) to eliminate phonetic stuttering and chopped acoustic artifacts. - Semantic Rhythm & Micro-Pauses: Dynamically calculates pause durations based on punctuation syntax (
.-> 220ms,...-> 370ms,—-> 320ms, dialogue turnarounds -> 400ms, paragraph transitions -> 650ms, chapter breaks -> 1,200ms).
- Pure Python Signal Processing: Custom 16-bit PCM waveform assembler with zero heavy audio library dependencies.
- Zero DC Offset & Pop Removal: Searches local waveform bounds for exact sign changes (zero-crossings) before splicing, eliminating speaker pops and transducer clicks.
- Equal-Power Crossfades: Applies smooth sinusoidal crossfading between consecutive speech bursts.
- Chunk Checkpointing: Resumable
.chunk_cacheallows long multi-hour novels to resume instantly without re-synthesizing rendered paragraphs.
- Manuscripts: Plain text (
.txt), Markdown (.mdwith YAML frontmatter), EPUB (.epub), PDF (.pdf), and HTML files. - Live Web Scraping: Ingests online fiction and articles directly from URLs with automated boilerplate removal, multi-table prose assembly, and anchor fragment targeting (
#Chapter2). - Multi-Chapter Splitting: Detects inline section headers and Roman numeral breaks, formatting each section as a discrete chapter track.
- Edge-TTS (
--backend edge) (Default): Free, zero-setup Microsoft Neural TTS voices (en-US-ChristopherNeural,en-US-GuyNeural,en-GB-SoniaNeural,en-US-JennyNeural). - Gemini / Cloud TTS (
--backend gemini): Google Gemini studio voices (Fenrir,Aoede,Puck,Charon,Kore,Leda,Oran,Zephyr). - OpenAI TTS (
--backend openai): High-definition OpenAI voices (tts-1-hdwithalloy,echo,fable,onyx,nova,shimmer). - Mock Engine (
--backend mock): Fast, offline synthetic harmonic engine for headless testing and CI/CD pipelines.
- M4B & MP3 Containers: Encodes to standard audiobook
.m4b(AAC) or.mp3with embedded ID3v2/MP4 chapter timepoint marks. - ITU-R BS.1770-4 / EBU R128 Normalization: Masters output to standard audiobook loudness targets (-16 LUFS, -1.5 dBFS true peak).
- Browser Studio UI (
parlando --web): Full-featured dark-mode console with live audio auditioning, progress bar telemetry, and cast inspection. - Real Web Audio FFT Visualizer: Real-time 40-bar frequency spectrum analyzer powered by HTML5
AudioContext+AnalyserNode. - Standalone Single-File Player: Automatically outputs a standalone, zero-dependency
<title>_player.htmlaudio player with interactive chapter scrubbing and speed controls.
git clone https://github.com/plbogen2/parlando.git
cd parlando
pip install -e .- Python:
3.9+ - System:
ffmpeg(recommended for M4B chapter embedding and broadcast loudness mastering)# Debian / Ubuntu sudo apt install ffmpeg # macOS brew install ffmpeg
Launch the browser studio on port 8765:
parlando --web --port 8765Open http://localhost:8765 to drag-and-drop manuscripts, paste text, scrape URLs, audition voices, and monitor live synthesis.
parlando novel.md \
-o audiobook.m4b \
--backend gemini \
--voice Fenrir \
-c "Case:male:Puck" \
-c "Linda Lee:female:Aoede" \
-c "Wintermute:neutral" \
--pacing dramaticparlando novel.md --dry-run -c "Case:male:Puck" -c "Linda Lee:female:Aoede"parlando https://www.infinityplus.co.uk/stories/colderwar.htm \
-o colder_war.m4b \
--voice en-US-GuyNeural \
--pacing cinematicparlando novel.epub --audition --voice en-GB-SoniaNeuralfrom parlando.pipeline import AudiobookPipeline, PipelineConfig
config = PipelineConfig(
backend="gemini",
voice="Fenrir", # Consistent Narrator voice
characters={
"Case": {"gender": "male", "voice": "Puck"},
"Linda Lee": {"gender": "female", "voice": "Aoede"},
},
pacing_mode="dramatic",
audio_format="mp3",
)
pipeline = AudiobookPipeline(config)
result = pipeline.run("novel.md", output_path="novel.mp3")
print(f"Master rendered: {result.audio_path} ({result.duration_seconds/60:.2f} min)")
for name, profile in result.characters.items():
print(f" - {name} ({profile.gender}): {result.voice_map[name]}")| Flag | Default | Description |
|---|---|---|
input |
(Required) | Filepath (.md, .txt, .epub, .pdf, .html) or web URL (https://...). |
-o, --output |
Auto-derived | Destination audio master path (.m4b, .mp3, .wav). |
--backend |
edge |
Synthesis engine: edge, openai, gemini, or mock. |
--voice |
en-US-ChristopherNeural |
Primary narrator neural voice name. |
--dialogue-voice |
None | Global voice override for character dialogue. |
-c, --character |
None | Character specification (e.g. 'Case:male:Puck', 'Linda:female', 'Clerk:male'). Repeatable. |
--cast |
None | Path to JSON file or JSON string defining character profiles and voice overrides. |
--pacing |
normal |
Pacing preset: normal, brisk, dramatic, cinematic, contemplative. |
--speed |
1.0 |
Global playback speed multiplier (0.5x – 2.0x). |
--format |
m4b |
Container format: m4b (chaptered AAC), mp3 (ID3v2), or wav. |
--crossfade |
35 |
Zero-crossing crossfade overlap duration in milliseconds. |
--audition |
False |
Auditioning mode: synthesizes only the first section/excerpt. |
--dry-run |
False |
Analyzes cast allocation, chapters, and word counts without synthesizing audio. |
--web |
False |
Launches the interactive browser studio and REST API daemon. |
--port |
8765 |
Port for the web studio server. |
Parlando can be run as a standalone container or as a companion sidecar for Marginalia:
docker build -t parlando .
docker run -p 8765:8765 parlandoParlando includes a comprehensive test suite covering DSP crossfades, contraction-safe parsing, speaker detection, scene-aware voice casting, and end-to-end pipeline execution:
python3 run_tests.pyParlando is open-source software licensed under the MIT License. Created by Dr. Paul Logasa Bogen II.