Frequently asked questions
What is synthetic speech used for in ML?
Synthetic speech accelerates machine learning training pipelines by providing consistent, watermark-marked audio corpora. LinguaForge delivers labeled audio datasets optimized for high-throughput generation, enabling developers to scale training data efficiently without the logistical constraints of human recording sessions.
How does 2Tiger differ from human-recorded datasets?
Unlike human-recorded datasets, our synthetic voices are generated on demand via API, ensuring consistent quality and scalability. Each file carries an inaudible watermark marking it as synthetic speech, and input text is not used for training, preserving data integrity for specific model architectures.
Are the voices licensed for commercial use?
Yes. Our voices are based on licensed recordings from LibriTTS-R (CC BY 4.0) and public-domain LibriVox audiobook readings. This licensing structure supports commercial deployment of generated audio corpora for machine learning training pipelines without additional royalty complexities for the end user.
Can I use the API to train my own voice models?
You can use the generated audio to train your own models, but we do not offer custom voice cloning. You cannot upload your own voice or design a custom voice. The API provides eight preset voices (e.g., Aria, Nova, Atlas) for consistent dataset generation.
What happens if a request fails?
If synthesis fails, the charge is returned in full. Failed requests are free, ensuring you only pay for successfully generated audio files. This protects your budget during high-throughput corpus generation where transient errors might occur under load.
Do you support SSML for finer control?
No, we do not support SSML. The API accepts simple text input up to 4096 characters. Control is limited to voice selection, speed (0.25–4.0), and language specification. For finer control, you must manage text segmentation and formatting in your application logic.
Is there a limit on characters per request?
Yes, each request accepts up to 4096 characters of input text. For longer texts, you must split the content into parts of this limit, generate audio for each part, and join the resulting audio files in your pipeline to create the final corpus.
How are spaces counted in pricing?
Spaces and punctuation count as characters for pricing. The rate is $10 per 1 million characters, or $0.01 per 1,000 characters. This means every space in your input text contributes to the total character count and associated cost.
What formats are best for ML training?
We support mp3, wav, opus, flac, and pcm. Select the format that best matches your model's input requirements. Select the format that best matches your model's input requirements.
Do you offer custom voice cloning?
No, we do not offer custom voice cloning for customers. You cannot upload your own voice or create custom voice designs. The API provides eight preset voices based on licensed recordings, ensuring consistent synthetic speech generation for your training datasets.
How long does synthesis take?
Typical speed is roughly real time, generating about 25 characters of text per second of audio. Short phrases take a few seconds. Requests are processed sequentially, so under load, a request may wait in a queue before synthesis begins.
Can I stream audio in real-time?
No, we do not support streaming or WebSocket connections. You send a POST request with text and receive the complete audio file in the response. This batch-oriented approach is optimized for generating large labeled audio corpora rather than real-time playback.
Is there a free trial available?
Yes, signing in with Google provides $0.50 of credit (about 50,000 characters) valid for 7 days. Accounts created with email and password start without trial credit. This allows you to test the API integration before committing to USDT top-ups.
How do I pay using USDT?
Top-ups are made via USDT on the TRON network (TRC-20) only. The minimum amount is $5. Credits are added automatically within minutes after the transfer is confirmed. There are no subscriptions or invoices; balance does not expire and remains available for future requests.
First request
curl
curl https://api.2tiger.info/v1/audio/speech \
-H "Authorization: Bearer $TTS_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"tts-1","voice":"aria","input":"Hello from the API.","response_format":"mp3"}' \
--output hello.mp3Engineered for Dataset Scalability
LinguaForge delivers consistent synthesis output, ensuring that identical inputs produce identical audio files. This consistency eliminates variance in training data, allowing machine learning pipelines to isolate acoustic features without noise from stochastic generation models. Each request generates a stable artifact suitable for large-scale corpus construction.
Deterministic Synthesis Output
The API provides consistent results for the same parameters, which is critical for benchmarking TTS models. By removing stochastic variability, developers can accurately measure improvements in their acoustic models against a consistent baseline.
- Identical inputs yield identical audio files
- Eliminates noise in acoustic feature analysis
- Stable baseline for model benchmarking
Consistent Voice Profiles
Eight preset voices maintain uniform tonal characteristics across all generated files. Whether using the warm narrator tone of Aria or the deep documentary style of Orion, each voice remains distinct and consistent, preventing profile drift in long-running dataset generation tasks.
- Uniform tonal characteristics across requests
- Distinct profiles for Aria, Nova, Luna, Sage, Felix, Milo, Atlas, Orion
- No voice drift during bulk generation
Batch-Ready API Structure
The API structure supports straightforward batching of up to 4096 characters per request. Clients can split larger texts and process them sequentially, receiving concatenated audio files that maintain temporal coherence. This design simplifies the ingestion of labeled corpora into standard ML workflows.
- Supports batch processing via sequential requests
- Max 4096 characters per input
- Concatenated output maintains temporal coherence
Licensed Voice Assets for Training Data
LinguaForge utilizes eight voice presets derived from the LibriTTS-R dataset (CC BY 4.0), ensuring consistent audio characteristics for reliable machine learning training. Each voice maintains stable tonal qualities across generations, which is critical for maintaining label integrity in large-scale corpora. The source material originates from LibriVox public-domain audiobook readings, providing a legally sound foundation for commercial synthetic speech production without complex licensing overhead.
All generated audio files contain an inaudible watermark that explicitly marks the content as synthetic speech. This feature allows downstream ML pipelines to easily distinguish between human-recorded and AI-generated data points during model training and evaluation phases.
LibriTTS-R Licensed Voices
The eight available voices are built on licensed recordings from the LibriTTS-R dataset. This ensures that the synthetic voices retain the acoustic properties of the original public-domain source material while offering distinct tonal profiles suitable for diverse dataset requirements.
- Female voices: aria, nova, luna, sage
- Male voices: felix, milo, atlas, orion
Public Domain Source Material
Source recordings come from LibriVox public-domain audiobook readings. This clear provenance simplifies compliance checks for data scientists building training sets, as the underlying audio assets are freely available for reuse and modification.
Inaudible Synthetic Watermarking
Every synthesized audio file includes a digital watermark indicating its synthetic origin. This metadata allows automated systems to filter or tag data points based on their generation method, improving dataset purity.
Supported Languages and Scripts
LinguaForge supports 23 languages, covering major global scripts including Arabic, Chinese, Cyrillic, and Latin-based systems. The API handles diverse phonetic structures without requiring separate voice models per language.
Our engine is optimized for high-throughput dataset generation, ensuring consistent audio quality across all supported tongues. This consistency is critical for training robust machine learning models on varied linguistic data.
23 Supported Languages
We support Arabic, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, and Turkish. Any voice preset can articulate any of these languages, allowing flexible voice selection regardless of linguistic origin.
- Arabic, Chinese, Danish, Dutch, English, Finnish, French, German, Greek, Hebrew, Hindi, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swahili, Swedish, Turkish
Automatic Language Detection
If the language field is omitted, the system guesses the language based on the input script. Text in Latin script defaults to English by default. This simplifies initial API calls for monolingual datasets.
Explicit Language Tagging
For non-English Latin-script text, explicitly set the language parameter to ensure correct pronunciation. This prevents misinterpretation of words that look similar but have different phonetic requirements in different languages.
Audio Formats for Model Ingestion
For most machine learning pipelines, MP3 provides an efficient balance between file size and audio quality. The API supports standard MP3 encoding as the default output, ensuring compatibility with widely used audio processing libraries. This format is suitable for large-scale corpus generation where storage costs and bandwidth efficiency are primary concerns.
Lossless formats such as FLAC and WAV are available for applications requiring bit-perfect audio reproduction. These formats preserve the full spectral detail of the synthetic speech, making them ideal for fine-grained acoustic feature extraction or high-fidelity model training. Select the appropriate format based on your downstream processing requirements.
Standard MP3 Encoding
MP3 is the recommended format for bulk dataset creation. It offers significant compression while maintaining intelligibility for speech recognition tasks. The API returns MP3 files by default, reducing transfer times and storage overhead for large character volumes.
Lossless FLAC and WAV
FLAC and WAV formats ensure no data loss during synthesis. These are critical for training models that rely on precise amplitude or frequency analysis. Use these formats when your algorithm requires exact signal representation without compression artifacts.
Raw PCM Stream Data
PCM output provides raw, uncompressed audio data. This format is ideal for custom audio buffers that handle their own encoding. It eliminates the need for post-processing decoding steps in your pipeline.
Developer-First Integration
LinguaForge integrates directly into existing ML pipelines using standard OpenAI SDKs. By updating the base URL to https://api.2tiger.info/v1, developers can utilize familiar client libraries for seamless API consumption without custom wrapper logic.
OpenAI SDK Compatibility
The API mirrors the OpenAI speech endpoint structure, allowing immediate adoption by teams already utilizing official OpenAI SDKs. This reduces integration friction and leverages established client-side codebases for reliable request handling and error management.
Simple JSON Payloads
Requests utilize a straightforward JSON body containing the text input, voice preset, and desired audio format. This standard structure ensures predictable parsing and easy scripting for automated dataset generation workflows.
- Supports standard fields: model, input, voice, response_format, and speed.
Instant Key Rotation
Security is maintained through single-key accounts with immediate revocation capabilities. Issuing a new key instantly invalidates the previous one, ensuring tight control over API access without waiting for session timeouts.

Transparent Pricing Model
Usage is billed strictly per character at $10 per 1 million characters, inclusive of spaces and punctuation. Failed requests incur no charge, ensuring you pay only for successful synthesis.
Per-Character Billing
Each character counts toward the total, including whitespace and punctuation marks. The rate remains constant regardless of voice selection or audio format chosen.
No Hidden Infrastructure Fees
The per-character rate covers all processing costs. There are no monthly subscriptions, seat fees, or additional charges for high-volume bursts or specific voice models.
Credit-Based Top-Ups
Add funds via USDT (TRC-20) with a $5 minimum. Credits are available automatically within minutes after the transfer. Balances do not expire and support unlimited usage.
Technical Constraints and Limits
The API enforces a strict 4096-character limit per request. For longer texts, split the input into chunks and concatenate the resulting audio files sequentially. Spaces and punctuation count toward this character total.
Requests are processed sequentially, meaning under load, synthesis may queue. Typical generation speed is roughly real time (about 25 characters of text per second of audio). Short phrases may take a few seconds to complete.
4096 Character Input Limit
Each API call accepts a maximum of 4096 characters in the input field. Exceeding this limit returns an error. For corpus generation, divide your dataset into segments that fit within this constraint and stitch the output audio files together programmatically.
- Limit applies to the raw character count, including spaces and punctuation.
- Split logic should preserve sentence boundaries to avoid audio artifacts.
- Concatenation is handled client-side; the API returns independent audio files.
Rate Limiting Architecture
Standard accounts are limited to 60 requests per minute. New accounts start with a 10 requests-per-minute limit until the first top-up is processed. This constraint ensures stable throughput for dataset generation pipelines.
Rate limits are applied per API key. If you exceed the limit, requests will be rejected with a 429 status code. Rotate keys or implement exponential backoff in your ingestion script to handle burst traffic.
- 60 requests/minute after initial top-up.
- 10 requests/minute for new accounts.
- Limits reset on a rolling one-minute window.
Queue Processing Times
Requests are processed in order. While typical generation speed is near real-time (25 chars/sec), high concurrency can introduce queue delays. Short phrases may take a few seconds to synthesize even without heavy load.
Failed requests incur no charge, but they still consume your rate limit allowance. Monitor your usage dashboard to track queue depth and adjust your submission frequency accordingly.
- Processing is sequential, not parallel.
- Queue depth varies based on concurrent requests.
- No SLA guarantees for maximum latency.
Use Cases in Synthetic Speech
LinguaForge enables precise control over synthetic speech generation, allowing developers to build robust training corpora for machine learning pipelines. By leveraging deterministic synthesis and consistent voice profiles, teams can generate labeled audio datasets that mirror real-world acoustic variations without the cost of human recording.
Domain Adaptation Data
Generate domain-specific audio by feeding distinct text corpora into the API. This allows for the creation of specialized training sets for verticals like finance, healthcare, or legal sectors, ensuring models learn domain-specific terminology and prosody patterns accurately.
- Generate text-specific audio for specialized verticals
- Ensure consistent prosody across labeled datasets
- Scale corpus creation without human recording bottlenecks
Accent and Dialect Variation
Leverage the 23 supported languages and 8 distinct voice presets to introduce acoustic diversity. While voices are consistent, the underlying engine handles multiple scripts and languages, allowing you to simulate regional variations by adjusting input text and language tags.
- Utilize 8 distinct voice presets for variety
- Support 23 languages with explicit tagging
- Simulate acoustic diversity for model robustness
Stress Testing TTS Models
Use the API to generate large volumes of synthetic audio for edge-case testing. The consistent watermark allows you to distinguish synthetic data from real recordings during model evaluation, ensuring your TTS model generalizes well across different input types.
- Generate high-volume audio for stress tests
- Use inaudible watermarks for data labeling
- Validate model performance on synthetic inputs
Getting Started with 2Tiger
Access the API key via the /get-api-key/ dashboard, where your unique credentials and current balance are displayed. The system supports instant key rotation; issuing a new token immediately invalidates the previous one, ensuring secure pipeline management for automated data collection.
Send a POST request to https://api.2tiger.info/v1/audio/speech with your text and voice parameters. The endpoint returns the audio file directly, allowing seamless integration into your data ingestion scripts without intermediate storage steps.
API Key Acquisition
Navigate to the /get-api-key/ page to generate your initial credentials. Signing in with Google grants a $0.50 trial credit (50,000 characters) valid for seven days, while email-based accounts start with a zero balance.
- Sign up via Google or email/password
- Retrieve your Bearer token immediately
- Rotate keys instantly if compromised
First Request Execution
Construct a JSON payload targeting https://api.2tiger.info/v1/audio/speech. Include the model identifier (tts-1 or tts-1-hd), the input text (max 4096 characters), and the desired voice preset. The API returns the synthesized audio in the specified format, such as MP3 or WAV, ready for immediate dataset labeling.
- POST to /v1/audio/speech
- Specify model, input, and voice
- Receive audio file in response
Monitoring Usage Metrics
The dashboard tracks character consumption and request status in real time. Failed requests are refunded automatically, ensuring you only pay for successful synthesis. Use the visible metrics to adjust batch sizes and maintain consistent throughput within the 60 requests-per-minute limit.
- Track character consumption
- Monitor request success rates
- Check balance before top-up
Listen to the 8 voices
Each sample below was generated by this API from the same sentence. Balance key issuance with throughput needs, as each key has a fixed rate limit.
23 languages
Any voice speaks any of these languages.
Sources and further reading
- Wikipedia
- LibriVox public-domain audiobooks
- curl documentation
- W3C WAI: making audio and video accessible
- ISO 639 language codes
- Node.js documentation
- FFmpeg documentation
- OpenAI API reference: create speech
- IETF RFC 7845: Ogg Opus
- LibriTTS-R corpus (OpenSLR 141)
- LibriTTS-R paper (arXiv 2305.18802)
- Tether (USDT)
