This is a simplified guide to an AI model called Autotune maintained by Nateraw. If you like these kinds of analysis, you should join AImodels.fyi or follow us on Twitter.
autotune
performs pitch correction on voice audio, applying auto-tuning effects to shift vocal frequencies to target pitches. Built by nateraw, this model processes audio input and returns corrected output in your chosen format. The core function is straightforward: it normalizes audio using one of three strategies (closest pitch, scale-based, or other methods) and outputs the pitch-corrected result. Before using this model, understand that it is a pitch-correction tool specifically designed for voice processing—it will not add effects, change vocal timbre substantially, or improve audio quality beyond pitch alignment. It accepts audio files via URI and outputs a single corrected audio file.
Vocal pitch correction for music production. Musicians and producers use autotune to align vocal recordings to a specific key or scale. If a vocalist drifts slightly sharp or flat during a take, autotune
can shift the pitch to match the intended notes without requiring a full re-recording. This is essential in modern music production where vocal precision is expected, especially for layered vocal arrangements or matching multiple vocal tracks.
Live performance pitch adjustment. Content creators and streamers processing recorded vocals in real-time workflows benefit from straightforward pitch correction. Unlike heavy autotune artifacts that announce their presence, this model allows you to make subtle or dramatic pitch shifts depending on the scale strategy chosen. This works particularly well for vocal tracks where the original performance is solid but pitch stability needs improvement.
Karaoke and cover track preparation. When preparing vocal tracks for karaoke or cover versions, pitch correction ensures the vocal stays locked to the backing track's key. If a singer naturally performs slightly sharp or flat, autotune
adjusts the vocal to perfect alignment without sounding robotic (depending on scale strategy selection).
Educational and speech clarity applications. Speech clarity sometimes improves when pitch variation is controlled. For educational content, instructional videos, or accessibility applications, slight pitch normalization can make speech more uniform and easier to process for both listeners and downstream speech recognition systems.
The model provides minimal documentation beyond its core function, making it difficult to understand advanced behavior or failure modes. It only processes single audio files per request—batch processing requires multiple API calls, which impacts efficiency for workflows with many audio files. The "scale" parameter determines how pitch correction applies, but the exact algorithms behind "closest," scale-based normalization, and other modes are undocumented; users must experiment to find which strategy suits their material.
Audio file size constraints are not specified in the schema, so you may encounter failures with very long recordings or large files without prior warning. The output format defaults to WAV, but if you need MP3 or other formats for distribution, you must ensure your downstream pipeline can handle format conversion. The model does not support pitch shifting to arbitrary target pitches—it only corrects to "natural" pitches based on the chosen scale, meaning fine-grained pitch control is limited.
Quality depends heavily on input audio characteristics. Clean, dry vocals process better than heavily compressed, distorted, or noisy recordings. The model will not remove background noise, improve tone, or add presence; it only addresses pitch. For low-quality source audio with pitch instability caused by poor microphone technique or background noise, results may be unsatisfactory. The model does not provide real-time processing, so live streaming applications require offline processing before broadcast.
speech-enhancer by playmore improves overall audio quality including noise reduction and clarity, whereas autotune
focuses only on pitch correction. Choose autotune
if pitch alignment is your primary goal; choose speech-enhancer if you need broader audio cleanup. The tradeoff is focus: autotune
does one thing precisely, while speech-enhancer handles multiple degradation problems.
qwenasr by twangodev performs speech recognition and alignment, detecting what words are spoken and their timing, while autotune
modifies only the pitch of audio without transcription. You would use autotune
for music production and voice modification; use qwenasr if you need to understand what was said or synchronize speech with text.
beat_this by xavriley detects rhythmic beats in music, solving an entirely different problem than pitch correction. These tools complement each other in a music production pipeline but serve separate functions; choose based on whether you need rhythm analysis or pitch adjustment.
music by elevenlabs composes and generates music from prompts or composition plans, creating new audio content from scratch. autotune
modifies existing voice recordings, so these models address opposite workflow needs: use music for creation, use autotune
for correction of recorded material.
v3 by elevenlabs is a text-to-speech model generating expressive synthetic speech, whereas autotune
processes human vocal recordings. Choose v3 if you need to generate speech from text; choose autotune
if you have recorded human voice that needs pitch correction.
The model runs on Replicate's infrastructure using Cog v0.10.0-alpha6+dev, a containerized framework for packaging machine learning models. It was last updated on 2024-04-19. The implementation uses a modular pitch-correction architecture that accepts audio via URI, applies pitch normalization based on a selected strategy, and returns corrected audio.
Architecture and processing:
Constraints and defaults:
import replicate
client = replicate.Client(api_token="your-replicate-api-token")
input_params = {
"audio_file": "https://example.com/your-vocal-recording.wav",
"scale": "closest", # Options: "closest" or other scale strategies
"output_format": "wav" # Output format: "wav", "mp3", "flac", etc.
}
output = client.run(
"nateraw/autotune:53d58aea27ccd949e5f9d77e4b2a74ffe90e1fa534295b257cea50f011e233dd",
input=input_params
)
print(f"Pitch-corrected audio: {output}")
import urllib.request
urllib.request.urlretrieve(output, "corrected_audio.wav")
Q: What audio formats does autotune accept as input?
A: The model accepts any audio format accessible via HTTP/HTTPS URI. Common formats include WAV, MP3, FLAC, and others. The audio must be hosted remotely and accessible via a direct URL; local file uploads are not supported—you must provide a publicly accessible URI.
Q: What does the "scale" parameter do, and which option should I choose?
A: The scale parameter determines the pitch-correction strategy. "Closest" (the default) normalizes each pitch to the nearest natural note, suitable for general vocal correction. Other scale-based strategies may enforce specific musical scales (major, minor, pentatonic, etc.), useful when you want the vocal locked to a particular key. Experiment with different options to find the sound that matches your music.
Q: How long does pitch correction take, and can I process multiple files at once?
A: Processing time is not documented in the model schema, but depends on audio duration and server load. The model processes one file per API request, so you cannot batch multiple files in a single call. For workflows with many audio files, you must submit separate API calls for each file or implement a queue to manage requests.
Q: Can I control the amount of pitch correction or make it more subtle?
A: The model does not expose a strength or intensity parameter. Pitch correction applies fully based on the chosen scale strategy—you cannot dial in a percentage or intensity level. If you need subtle vs. aggressive correction options, you may need to experiment with different scale strategies or use alternative tools with more granular control.
Q: What output formats are available, and can I choose formats other than WAV?
A: The output_format parameter supports multiple formats including WAV (default), MP3, FLAC, and others. Choose based on your distribution needs—WAV and FLAC preserve quality for archival or further processing, while MP3 reduces file size for streaming or download. The default WAV format is lossless and suitable for production use.
Q: Is autotune suitable for production music or only demo/testing?
A: The model is suitable for production music production workflows where pitch correction is needed. However, success depends on input quality and the scale strategy chosen. Professional results require clean vocal recordings and careful selection of the appropriate scale to match your musical key. The lack of strength/intensity controls may limit flexibility compared to dedicated DAW autotune plugins.
Q: How does the output file URI work, and how long can I access the corrected audio?
A: The model returns a URI string pointing to the corrected audio hosted on Replicate's infrastructure. The file remains accessible for a limited period (typically 24-48 hours) before automatic cleanup. Download or store the corrected audio immediately if you plan to use it beyond this window.
Q: What happens if I submit a very long audio file, like a full song with multiple verses?
A: Audio file size and duration limits are not explicitly specified. Very long files may timeout or fail, but exact limits are undocumented. Test with your typical audio durations to determine feasibility; if processing fails silently or times out, try breaking longer recordings into shorter segments.